1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/memnode.hpp" 39 #include "opto/mulnode.hpp" 40 #include "opto/narrowptrnode.hpp" 41 #include "opto/phaseX.hpp" 42 #include "opto/regmask.hpp" 43 #include "utilities/copy.hpp" 44 45 // Portions of code courtesy of Clifford Click 46 47 // Optimization - Graph Style 48 49 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); 50 51 //============================================================================= 52 uint MemNode::size_of() const { return sizeof(*this); } 53 54 const TypePtr *MemNode::adr_type() const { 55 Node* adr = in(Address); 56 const TypePtr* cross_check = NULL; 57 DEBUG_ONLY(cross_check = _adr_type); 58 return calculate_adr_type(adr->bottom_type(), cross_check); 59 } 60 61 #ifndef PRODUCT 62 void MemNode::dump_spec(outputStream *st) const { 63 if (in(Address) == NULL) return; // node is dead 64 #ifndef ASSERT 65 // fake the missing field 66 const TypePtr* _adr_type = NULL; 67 if (in(Address) != NULL) 68 _adr_type = in(Address)->bottom_type()->isa_ptr(); 69 #endif 70 dump_adr_type(this, _adr_type, st); 71 72 Compile* C = Compile::current(); 73 if( C->alias_type(_adr_type)->is_volatile() ) 74 st->print(" Volatile!"); 75 } 76 77 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { 78 st->print(" @"); 79 if (adr_type == NULL) { 80 st->print("NULL"); 81 } else { 82 adr_type->dump_on(st); 83 Compile* C = Compile::current(); 84 Compile::AliasType* atp = NULL; 85 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); 86 if (atp == NULL) 87 st->print(", idx=?\?;"); 88 else if (atp->index() == Compile::AliasIdxBot) 89 st->print(", idx=Bot;"); 90 else if (atp->index() == Compile::AliasIdxTop) 91 st->print(", idx=Top;"); 92 else if (atp->index() == Compile::AliasIdxRaw) 93 st->print(", idx=Raw;"); 94 else { 95 ciField* field = atp->field(); 96 if (field) { 97 st->print(", name="); 98 field->print_name_on(st); 99 } 100 st->print(", idx=%d;", atp->index()); 101 } 102 } 103 } 104 105 extern void print_alias_types(); 106 107 #endif 108 109 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { 110 assert((t_oop != NULL), "sanity"); 111 bool is_instance = t_oop->is_known_instance_field(); 112 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && 113 (load != NULL) && load->is_Load() && 114 (phase->is_IterGVN() != NULL); 115 if (!(is_instance || is_boxed_value_load)) 116 return mchain; // don't try to optimize non-instance types 117 uint instance_id = t_oop->instance_id(); 118 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); 119 Node *prev = NULL; 120 Node *result = mchain; 121 while (prev != result) { 122 prev = result; 123 if (result == start_mem) 124 break; // hit one of our sentinels 125 // skip over a call which does not affect this memory slice 126 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { 127 Node *proj_in = result->in(0); 128 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { 129 break; // hit one of our sentinels 130 } else if (proj_in->is_Call()) { 131 CallNode *call = proj_in->as_Call(); 132 if (!call->may_modify(t_oop, phase)) { // returns false for instances 133 result = call->in(TypeFunc::Memory); 134 } 135 } else if (proj_in->is_Initialize()) { 136 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); 137 // Stop if this is the initialization for the object instance which 138 // which contains this memory slice, otherwise skip over it. 139 if ((alloc == NULL) || (alloc->_idx == instance_id)) { 140 break; 141 } 142 if (is_instance) { 143 result = proj_in->in(TypeFunc::Memory); 144 } else if (is_boxed_value_load) { 145 Node* klass = alloc->in(AllocateNode::KlassNode); 146 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); 147 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { 148 result = proj_in->in(TypeFunc::Memory); // not related allocation 149 } 150 } 151 } else if (proj_in->is_MemBar()) { 152 result = proj_in->in(TypeFunc::Memory); 153 } else { 154 assert(false, "unexpected projection"); 155 } 156 } else if (result->is_ClearArray()) { 157 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { 158 // Can not bypass initialization of the instance 159 // we are looking for. 160 break; 161 } 162 // Otherwise skip it (the call updated 'result' value). 163 } else if (result->is_MergeMem()) { 164 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); 165 } 166 } 167 return result; 168 } 169 170 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { 171 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); 172 if (t_oop == NULL) 173 return mchain; // don't try to optimize non-oop types 174 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); 175 bool is_instance = t_oop->is_known_instance_field(); 176 PhaseIterGVN *igvn = phase->is_IterGVN(); 177 if (is_instance && igvn != NULL && result->is_Phi()) { 178 PhiNode *mphi = result->as_Phi(); 179 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); 180 const TypePtr *t = mphi->adr_type(); 181 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || 182 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && 183 t->is_oopptr()->cast_to_exactness(true) 184 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) 185 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { 186 // clone the Phi with our address type 187 result = mphi->split_out_instance(t_adr, igvn); 188 } else { 189 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); 190 } 191 } 192 return result; 193 } 194 195 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { 196 uint alias_idx = phase->C->get_alias_index(tp); 197 Node *mem = mmem; 198 #ifdef ASSERT 199 { 200 // Check that current type is consistent with the alias index used during graph construction 201 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); 202 bool consistent = adr_check == NULL || adr_check->empty() || 203 phase->C->must_alias(adr_check, alias_idx ); 204 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] 205 if( !consistent && adr_check != NULL && !adr_check->empty() && 206 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && 207 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && 208 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || 209 adr_check->offset() == oopDesc::klass_offset_in_bytes() || 210 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { 211 // don't assert if it is dead code. 212 consistent = true; 213 } 214 if( !consistent ) { 215 st->print("alias_idx==%d, adr_check==", alias_idx); 216 if( adr_check == NULL ) { 217 st->print("NULL"); 218 } else { 219 adr_check->dump(); 220 } 221 st->cr(); 222 print_alias_types(); 223 assert(consistent, "adr_check must match alias idx"); 224 } 225 } 226 #endif 227 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally 228 // means an array I have not precisely typed yet. Do not do any 229 // alias stuff with it any time soon. 230 const TypeOopPtr *toop = tp->isa_oopptr(); 231 if( tp->base() != Type::AnyPtr && 232 !(toop && 233 toop->klass() != NULL && 234 toop->klass()->is_java_lang_Object() && 235 toop->offset() == Type::OffsetBot) ) { 236 // compress paths and change unreachable cycles to TOP 237 // If not, we can update the input infinitely along a MergeMem cycle 238 // Equivalent code in PhiNode::Ideal 239 Node* m = phase->transform(mmem); 240 // If transformed to a MergeMem, get the desired slice 241 // Otherwise the returned node represents memory for every slice 242 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; 243 // Update input if it is progress over what we have now 244 } 245 return mem; 246 } 247 248 //--------------------------Ideal_common--------------------------------------- 249 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. 250 // Unhook non-raw memories from complete (macro-expanded) initializations. 251 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 252 // If our control input is a dead region, kill all below the region 253 Node *ctl = in(MemNode::Control); 254 if (ctl && remove_dead_region(phase, can_reshape)) 255 return this; 256 ctl = in(MemNode::Control); 257 // Don't bother trying to transform a dead node 258 if (ctl && ctl->is_top()) return NodeSentinel; 259 260 PhaseIterGVN *igvn = phase->is_IterGVN(); 261 // Wait if control on the worklist. 262 if (ctl && can_reshape && igvn != NULL) { 263 Node* bol = NULL; 264 Node* cmp = NULL; 265 if (ctl->in(0)->is_If()) { 266 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); 267 bol = ctl->in(0)->in(1); 268 if (bol->is_Bool()) 269 cmp = ctl->in(0)->in(1)->in(1); 270 } 271 if (igvn->_worklist.member(ctl) || 272 (bol != NULL && igvn->_worklist.member(bol)) || 273 (cmp != NULL && igvn->_worklist.member(cmp)) ) { 274 // This control path may be dead. 275 // Delay this memory node transformation until the control is processed. 276 phase->is_IterGVN()->_worklist.push(this); 277 return NodeSentinel; // caller will return NULL 278 } 279 } 280 // Ignore if memory is dead, or self-loop 281 Node *mem = in(MemNode::Memory); 282 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL 283 assert(mem != this, "dead loop in MemNode::Ideal"); 284 285 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { 286 // This memory slice may be dead. 287 // Delay this mem node transformation until the memory is processed. 288 phase->is_IterGVN()->_worklist.push(this); 289 return NodeSentinel; // caller will return NULL 290 } 291 292 Node *address = in(MemNode::Address); 293 const Type *t_adr = phase->type(address); 294 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL 295 296 if (can_reshape && igvn != NULL && 297 (igvn->_worklist.member(address) || 298 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { 299 // The address's base and type may change when the address is processed. 300 // Delay this mem node transformation until the address is processed. 301 phase->is_IterGVN()->_worklist.push(this); 302 return NodeSentinel; // caller will return NULL 303 } 304 305 // Do NOT remove or optimize the next lines: ensure a new alias index 306 // is allocated for an oop pointer type before Escape Analysis. 307 // Note: C++ will not remove it since the call has side effect. 308 if (t_adr->isa_oopptr()) { 309 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); 310 } 311 312 Node* base = NULL; 313 if (address->is_AddP()) { 314 base = address->in(AddPNode::Base); 315 } 316 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && 317 !t_adr->isa_rawptr()) { 318 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. 319 // Skip this node optimization if its address has TOP base. 320 return NodeSentinel; // caller will return NULL 321 } 322 323 // Avoid independent memory operations 324 Node* old_mem = mem; 325 326 // The code which unhooks non-raw memories from complete (macro-expanded) 327 // initializations was removed. After macro-expansion all stores catched 328 // by Initialize node became raw stores and there is no information 329 // which memory slices they modify. So it is unsafe to move any memory 330 // operation above these stores. Also in most cases hooked non-raw memories 331 // were already unhooked by using information from detect_ptr_independence() 332 // and find_previous_store(). 333 334 if (mem->is_MergeMem()) { 335 MergeMemNode* mmem = mem->as_MergeMem(); 336 const TypePtr *tp = t_adr->is_ptr(); 337 338 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); 339 } 340 341 if (mem != old_mem) { 342 set_req(MemNode::Memory, mem); 343 if (can_reshape && old_mem->outcnt() == 0) { 344 igvn->_worklist.push(old_mem); 345 } 346 if (phase->type( mem ) == Type::TOP) return NodeSentinel; 347 return this; 348 } 349 350 // let the subclass continue analyzing... 351 return NULL; 352 } 353 354 // Helper function for proving some simple control dominations. 355 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. 356 // Already assumes that 'dom' is available at 'sub', and that 'sub' 357 // is not a constant (dominated by the method's StartNode). 358 // Used by MemNode::find_previous_store to prove that the 359 // control input of a memory operation predates (dominates) 360 // an allocation it wants to look past. 361 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { 362 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) 363 return false; // Conservative answer for dead code 364 365 // Check 'dom'. Skip Proj and CatchProj nodes. 366 dom = dom->find_exact_control(dom); 367 if (dom == NULL || dom->is_top()) 368 return false; // Conservative answer for dead code 369 370 if (dom == sub) { 371 // For the case when, for example, 'sub' is Initialize and the original 372 // 'dom' is Proj node of the 'sub'. 373 return false; 374 } 375 376 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) 377 return true; 378 379 // 'dom' dominates 'sub' if its control edge and control edges 380 // of all its inputs dominate or equal to sub's control edge. 381 382 // Currently 'sub' is either Allocate, Initialize or Start nodes. 383 // Or Region for the check in LoadNode::Ideal(); 384 // 'sub' should have sub->in(0) != NULL. 385 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || 386 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); 387 388 // Get control edge of 'sub'. 389 Node* orig_sub = sub; 390 sub = sub->find_exact_control(sub->in(0)); 391 if (sub == NULL || sub->is_top()) 392 return false; // Conservative answer for dead code 393 394 assert(sub->is_CFG(), "expecting control"); 395 396 if (sub == dom) 397 return true; 398 399 if (sub->is_Start() || sub->is_Root()) 400 return false; 401 402 { 403 // Check all control edges of 'dom'. 404 405 ResourceMark rm; 406 Arena* arena = Thread::current()->resource_area(); 407 Node_List nlist(arena); 408 Unique_Node_List dom_list(arena); 409 410 dom_list.push(dom); 411 bool only_dominating_controls = false; 412 413 for (uint next = 0; next < dom_list.size(); next++) { 414 Node* n = dom_list.at(next); 415 if (n == orig_sub) 416 return false; // One of dom's inputs dominated by sub. 417 if (!n->is_CFG() && n->pinned()) { 418 // Check only own control edge for pinned non-control nodes. 419 n = n->find_exact_control(n->in(0)); 420 if (n == NULL || n->is_top()) 421 return false; // Conservative answer for dead code 422 assert(n->is_CFG(), "expecting control"); 423 dom_list.push(n); 424 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { 425 only_dominating_controls = true; 426 } else if (n->is_CFG()) { 427 if (n->dominates(sub, nlist)) 428 only_dominating_controls = true; 429 else 430 return false; 431 } else { 432 // First, own control edge. 433 Node* m = n->find_exact_control(n->in(0)); 434 if (m != NULL) { 435 if (m->is_top()) 436 return false; // Conservative answer for dead code 437 dom_list.push(m); 438 } 439 // Now, the rest of edges. 440 uint cnt = n->req(); 441 for (uint i = 1; i < cnt; i++) { 442 m = n->find_exact_control(n->in(i)); 443 if (m == NULL || m->is_top()) 444 continue; 445 dom_list.push(m); 446 } 447 } 448 } 449 return only_dominating_controls; 450 } 451 } 452 453 //---------------------detect_ptr_independence--------------------------------- 454 // Used by MemNode::find_previous_store to prove that two base 455 // pointers are never equal. 456 // The pointers are accompanied by their associated allocations, 457 // if any, which have been previously discovered by the caller. 458 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, 459 Node* p2, AllocateNode* a2, 460 PhaseTransform* phase) { 461 // Attempt to prove that these two pointers cannot be aliased. 462 // They may both manifestly be allocations, and they should differ. 463 // Or, if they are not both allocations, they can be distinct constants. 464 // Otherwise, one is an allocation and the other a pre-existing value. 465 if (a1 == NULL && a2 == NULL) { // neither an allocation 466 return (p1 != p2) && p1->is_Con() && p2->is_Con(); 467 } else if (a1 != NULL && a2 != NULL) { // both allocations 468 return (a1 != a2); 469 } else if (a1 != NULL) { // one allocation a1 470 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) 471 return all_controls_dominate(p2, a1); 472 } else { //(a2 != NULL) // one allocation a2 473 return all_controls_dominate(p1, a2); 474 } 475 return false; 476 } 477 478 479 // The logic for reordering loads and stores uses four steps: 480 // (a) Walk carefully past stores and initializations which we 481 // can prove are independent of this load. 482 // (b) Observe that the next memory state makes an exact match 483 // with self (load or store), and locate the relevant store. 484 // (c) Ensure that, if we were to wire self directly to the store, 485 // the optimizer would fold it up somehow. 486 // (d) Do the rewiring, and return, depending on some other part of 487 // the optimizer to fold up the load. 488 // This routine handles steps (a) and (b). Steps (c) and (d) are 489 // specific to loads and stores, so they are handled by the callers. 490 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) 491 // 492 Node* MemNode::find_previous_store(PhaseTransform* phase) { 493 Node* ctrl = in(MemNode::Control); 494 Node* adr = in(MemNode::Address); 495 intptr_t offset = 0; 496 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 497 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 498 499 if (offset == Type::OffsetBot) 500 return NULL; // cannot unalias unless there are precise offsets 501 502 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); 503 504 intptr_t size_in_bytes = memory_size(); 505 506 Node* mem = in(MemNode::Memory); // start searching here... 507 508 int cnt = 50; // Cycle limiter 509 for (;;) { // While we can dance past unrelated stores... 510 if (--cnt < 0) break; // Caught in cycle or a complicated dance? 511 512 if (mem->is_Store()) { 513 Node* st_adr = mem->in(MemNode::Address); 514 intptr_t st_offset = 0; 515 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 516 if (st_base == NULL) 517 break; // inscrutable pointer 518 if (st_offset != offset && st_offset != Type::OffsetBot) { 519 const int MAX_STORE = BytesPerLong; 520 if (st_offset >= offset + size_in_bytes || 521 st_offset <= offset - MAX_STORE || 522 st_offset <= offset - mem->as_Store()->memory_size()) { 523 // Success: The offsets are provably independent. 524 // (You may ask, why not just test st_offset != offset and be done? 525 // The answer is that stores of different sizes can co-exist 526 // in the same sequence of RawMem effects. We sometimes initialize 527 // a whole 'tile' of array elements with a single jint or jlong.) 528 mem = mem->in(MemNode::Memory); 529 continue; // (a) advance through independent store memory 530 } 531 } 532 if (st_base != base && 533 detect_ptr_independence(base, alloc, 534 st_base, 535 AllocateNode::Ideal_allocation(st_base, phase), 536 phase)) { 537 // Success: The bases are provably independent. 538 mem = mem->in(MemNode::Memory); 539 continue; // (a) advance through independent store memory 540 } 541 542 // (b) At this point, if the bases or offsets do not agree, we lose, 543 // since we have not managed to prove 'this' and 'mem' independent. 544 if (st_base == base && st_offset == offset) { 545 return mem; // let caller handle steps (c), (d) 546 } 547 548 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 549 InitializeNode* st_init = mem->in(0)->as_Initialize(); 550 AllocateNode* st_alloc = st_init->allocation(); 551 if (st_alloc == NULL) 552 break; // something degenerated 553 bool known_identical = false; 554 bool known_independent = false; 555 if (alloc == st_alloc) 556 known_identical = true; 557 else if (alloc != NULL) 558 known_independent = true; 559 else if (all_controls_dominate(this, st_alloc)) 560 known_independent = true; 561 562 if (known_independent) { 563 // The bases are provably independent: Either they are 564 // manifestly distinct allocations, or else the control 565 // of this load dominates the store's allocation. 566 int alias_idx = phase->C->get_alias_index(adr_type()); 567 if (alias_idx == Compile::AliasIdxRaw) { 568 mem = st_alloc->in(TypeFunc::Memory); 569 } else { 570 mem = st_init->memory(alias_idx); 571 } 572 continue; // (a) advance through independent store memory 573 } 574 575 // (b) at this point, if we are not looking at a store initializing 576 // the same allocation we are loading from, we lose. 577 if (known_identical) { 578 // From caller, can_see_stored_value will consult find_captured_store. 579 return mem; // let caller handle steps (c), (d) 580 } 581 582 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { 583 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. 584 if (mem->is_Proj() && mem->in(0)->is_Call()) { 585 CallNode *call = mem->in(0)->as_Call(); 586 if (!call->may_modify(addr_t, phase)) { 587 mem = call->in(TypeFunc::Memory); 588 continue; // (a) advance through independent call memory 589 } 590 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { 591 mem = mem->in(0)->in(TypeFunc::Memory); 592 continue; // (a) advance through independent MemBar memory 593 } else if (mem->is_ClearArray()) { 594 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { 595 // (the call updated 'mem' value) 596 continue; // (a) advance through independent allocation memory 597 } else { 598 // Can not bypass initialization of the instance 599 // we are looking for. 600 return mem; 601 } 602 } else if (mem->is_MergeMem()) { 603 int alias_idx = phase->C->get_alias_index(adr_type()); 604 mem = mem->as_MergeMem()->memory_at(alias_idx); 605 continue; // (a) advance through independent MergeMem memory 606 } 607 } 608 609 // Unless there is an explicit 'continue', we must bail out here, 610 // because 'mem' is an inscrutable memory state (e.g., a call). 611 break; 612 } 613 614 return NULL; // bail out 615 } 616 617 //----------------------calculate_adr_type------------------------------------- 618 // Helper function. Notices when the given type of address hits top or bottom. 619 // Also, asserts a cross-check of the type against the expected address type. 620 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { 621 if (t == Type::TOP) return NULL; // does not touch memory any more? 622 #ifdef PRODUCT 623 cross_check = NULL; 624 #else 625 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; 626 #endif 627 const TypePtr* tp = t->isa_ptr(); 628 if (tp == NULL) { 629 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); 630 return TypePtr::BOTTOM; // touches lots of memory 631 } else { 632 #ifdef ASSERT 633 // %%%% [phh] We don't check the alias index if cross_check is 634 // TypeRawPtr::BOTTOM. Needs to be investigated. 635 if (cross_check != NULL && 636 cross_check != TypePtr::BOTTOM && 637 cross_check != TypeRawPtr::BOTTOM) { 638 // Recheck the alias index, to see if it has changed (due to a bug). 639 Compile* C = Compile::current(); 640 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), 641 "must stay in the original alias category"); 642 // The type of the address must be contained in the adr_type, 643 // disregarding "null"-ness. 644 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) 645 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); 646 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), 647 "real address must not escape from expected memory type"); 648 } 649 #endif 650 return tp; 651 } 652 } 653 654 //------------------------adr_phi_is_loop_invariant---------------------------- 655 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted 656 // loop is loop invariant. Make a quick traversal of Phi and associated 657 // CastPP nodes, looking to see if they are a closed group within the loop. 658 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { 659 // The idea is that the phi-nest must boil down to only CastPP nodes 660 // with the same data. This implies that any path into the loop already 661 // includes such a CastPP, and so the original cast, whatever its input, 662 // must be covered by an equivalent cast, with an earlier control input. 663 ResourceMark rm; 664 665 // The loop entry input of the phi should be the unique dominating 666 // node for every Phi/CastPP in the loop. 667 Unique_Node_List closure; 668 closure.push(adr_phi->in(LoopNode::EntryControl)); 669 670 // Add the phi node and the cast to the worklist. 671 Unique_Node_List worklist; 672 worklist.push(adr_phi); 673 if( cast != NULL ){ 674 if( !cast->is_ConstraintCast() ) return false; 675 worklist.push(cast); 676 } 677 678 // Begin recursive walk of phi nodes. 679 while( worklist.size() ){ 680 // Take a node off the worklist 681 Node *n = worklist.pop(); 682 if( !closure.member(n) ){ 683 // Add it to the closure. 684 closure.push(n); 685 // Make a sanity check to ensure we don't waste too much time here. 686 if( closure.size() > 20) return false; 687 // This node is OK if: 688 // - it is a cast of an identical value 689 // - or it is a phi node (then we add its inputs to the worklist) 690 // Otherwise, the node is not OK, and we presume the cast is not invariant 691 if( n->is_ConstraintCast() ){ 692 worklist.push(n->in(1)); 693 } else if( n->is_Phi() ) { 694 for( uint i = 1; i < n->req(); i++ ) { 695 worklist.push(n->in(i)); 696 } 697 } else { 698 return false; 699 } 700 } 701 } 702 703 // Quit when the worklist is empty, and we've found no offending nodes. 704 return true; 705 } 706 707 //------------------------------Ideal_DU_postCCP------------------------------- 708 // Find any cast-away of null-ness and keep its control. Null cast-aways are 709 // going away in this pass and we need to make this memory op depend on the 710 // gating null check. 711 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { 712 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); 713 } 714 715 // I tried to leave the CastPP's in. This makes the graph more accurate in 716 // some sense; we get to keep around the knowledge that an oop is not-null 717 // after some test. Alas, the CastPP's interfere with GVN (some values are 718 // the regular oop, some are the CastPP of the oop, all merge at Phi's which 719 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed 720 // some of the more trivial cases in the optimizer. Removing more useless 721 // Phi's started allowing Loads to illegally float above null checks. I gave 722 // up on this approach. CNC 10/20/2000 723 // This static method may be called not from MemNode (EncodePNode calls it). 724 // Only the control edge of the node 'n' might be updated. 725 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { 726 Node *skipped_cast = NULL; 727 // Need a null check? Regular static accesses do not because they are 728 // from constant addresses. Array ops are gated by the range check (which 729 // always includes a NULL check). Just check field ops. 730 if( n->in(MemNode::Control) == NULL ) { 731 // Scan upwards for the highest location we can place this memory op. 732 while( true ) { 733 switch( adr->Opcode() ) { 734 735 case Op_AddP: // No change to NULL-ness, so peek thru AddP's 736 adr = adr->in(AddPNode::Base); 737 continue; 738 739 case Op_DecodeN: // No change to NULL-ness, so peek thru 740 case Op_DecodeNKlass: 741 adr = adr->in(1); 742 continue; 743 744 case Op_EncodeP: 745 case Op_EncodePKlass: 746 // EncodeP node's control edge could be set by this method 747 // when EncodeP node depends on CastPP node. 748 // 749 // Use its control edge for memory op because EncodeP may go away 750 // later when it is folded with following or preceding DecodeN node. 751 if (adr->in(0) == NULL) { 752 // Keep looking for cast nodes. 753 adr = adr->in(1); 754 continue; 755 } 756 ccp->hash_delete(n); 757 n->set_req(MemNode::Control, adr->in(0)); 758 ccp->hash_insert(n); 759 return n; 760 761 case Op_CastPP: 762 // If the CastPP is useless, just peek on through it. 763 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { 764 // Remember the cast that we've peeked though. If we peek 765 // through more than one, then we end up remembering the highest 766 // one, that is, if in a loop, the one closest to the top. 767 skipped_cast = adr; 768 adr = adr->in(1); 769 continue; 770 } 771 // CastPP is going away in this pass! We need this memory op to be 772 // control-dependent on the test that is guarding the CastPP. 773 ccp->hash_delete(n); 774 n->set_req(MemNode::Control, adr->in(0)); 775 ccp->hash_insert(n); 776 return n; 777 778 case Op_Phi: 779 // Attempt to float above a Phi to some dominating point. 780 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { 781 // If we've already peeked through a Cast (which could have set the 782 // control), we can't float above a Phi, because the skipped Cast 783 // may not be loop invariant. 784 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { 785 adr = adr->in(1); 786 continue; 787 } 788 } 789 790 // Intentional fallthrough! 791 792 // No obvious dominating point. The mem op is pinned below the Phi 793 // by the Phi itself. If the Phi goes away (no true value is merged) 794 // then the mem op can float, but not indefinitely. It must be pinned 795 // behind the controls leading to the Phi. 796 case Op_CheckCastPP: 797 // These usually stick around to change address type, however a 798 // useless one can be elided and we still need to pick up a control edge 799 if (adr->in(0) == NULL) { 800 // This CheckCastPP node has NO control and is likely useless. But we 801 // need check further up the ancestor chain for a control input to keep 802 // the node in place. 4959717. 803 skipped_cast = adr; 804 adr = adr->in(1); 805 continue; 806 } 807 ccp->hash_delete(n); 808 n->set_req(MemNode::Control, adr->in(0)); 809 ccp->hash_insert(n); 810 return n; 811 812 // List of "safe" opcodes; those that implicitly block the memory 813 // op below any null check. 814 case Op_CastX2P: // no null checks on native pointers 815 case Op_Parm: // 'this' pointer is not null 816 case Op_LoadP: // Loading from within a klass 817 case Op_LoadN: // Loading from within a klass 818 case Op_LoadKlass: // Loading from within a klass 819 case Op_LoadNKlass: // Loading from within a klass 820 case Op_ConP: // Loading from a klass 821 case Op_ConN: // Loading from a klass 822 case Op_ConNKlass: // Loading from a klass 823 case Op_CreateEx: // Sucking up the guts of an exception oop 824 case Op_Con: // Reading from TLS 825 case Op_CMoveP: // CMoveP is pinned 826 case Op_CMoveN: // CMoveN is pinned 827 break; // No progress 828 829 case Op_Proj: // Direct call to an allocation routine 830 case Op_SCMemProj: // Memory state from store conditional ops 831 #ifdef ASSERT 832 { 833 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); 834 const Node* call = adr->in(0); 835 if (call->is_CallJava()) { 836 const CallJavaNode* call_java = call->as_CallJava(); 837 const TypeTuple *r = call_java->tf()->range(); 838 assert(r->cnt() > TypeFunc::Parms, "must return value"); 839 const Type* ret_type = r->field_at(TypeFunc::Parms); 840 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); 841 // We further presume that this is one of 842 // new_instance_Java, new_array_Java, or 843 // the like, but do not assert for this. 844 } else if (call->is_Allocate()) { 845 // similar case to new_instance_Java, etc. 846 } else if (!call->is_CallLeaf()) { 847 // Projections from fetch_oop (OSR) are allowed as well. 848 ShouldNotReachHere(); 849 } 850 } 851 #endif 852 break; 853 default: 854 ShouldNotReachHere(); 855 } 856 break; 857 } 858 } 859 860 return NULL; // No progress 861 } 862 863 864 //============================================================================= 865 uint LoadNode::size_of() const { return sizeof(*this); } 866 uint LoadNode::cmp( const Node &n ) const 867 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } 868 const Type *LoadNode::bottom_type() const { return _type; } 869 uint LoadNode::ideal_reg() const { 870 return _type->ideal_reg(); 871 } 872 873 #ifndef PRODUCT 874 void LoadNode::dump_spec(outputStream *st) const { 875 MemNode::dump_spec(st); 876 if( !Verbose && !WizardMode ) { 877 // standard dump does this in Verbose and WizardMode 878 st->print(" #"); _type->dump_on(st); 879 } 880 } 881 #endif 882 883 #ifdef ASSERT 884 //----------------------------is_immutable_value------------------------------- 885 // Helper function to allow a raw load without control edge for some cases 886 bool LoadNode::is_immutable_value(Node* adr) { 887 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && 888 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 889 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == 890 in_bytes(JavaThread::osthread_offset()))); 891 } 892 #endif 893 894 //----------------------------LoadNode::make----------------------------------- 895 // Polymorphic factory method: 896 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) { 897 Compile* C = gvn.C; 898 899 // sanity check the alias category against the created node type 900 assert(!(adr_type->isa_oopptr() && 901 adr_type->offset() == oopDesc::klass_offset_in_bytes()), 902 "use LoadKlassNode instead"); 903 assert(!(adr_type->isa_aryptr() && 904 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), 905 "use LoadRangeNode instead"); 906 // Check control edge of raw loads 907 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 908 // oop will be recorded in oop map if load crosses safepoint 909 rt->isa_oopptr() || is_immutable_value(adr), 910 "raw memory operations should have control edge"); 911 switch (bt) { 912 case T_BOOLEAN: return new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo); 913 case T_BYTE: return new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo); 914 case T_INT: return new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo); 915 case T_CHAR: return new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo); 916 case T_SHORT: return new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo); 917 case T_LONG: return new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo); 918 case T_FLOAT: return new LoadFNode (ctl, mem, adr, adr_type, rt, mo); 919 case T_DOUBLE: return new LoadDNode (ctl, mem, adr, adr_type, rt, mo); 920 case T_ADDRESS: return new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo); 921 case T_OBJECT: 922 #ifdef _LP64 923 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 924 Node* load = gvn.transform(new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo)); 925 return new DecodeNNode(load, load->bottom_type()->make_ptr()); 926 } else 927 #endif 928 { 929 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); 930 return new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo); 931 } 932 } 933 ShouldNotReachHere(); 934 return (LoadNode*)NULL; 935 } 936 937 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) { 938 bool require_atomic = true; 939 return new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic); 940 } 941 942 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) { 943 bool require_atomic = true; 944 return new LoadDNode(ctl, mem, adr, adr_type, rt, mo, require_atomic); 945 } 946 947 948 949 //------------------------------hash------------------------------------------- 950 uint LoadNode::hash() const { 951 // unroll addition of interesting fields 952 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); 953 } 954 955 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { 956 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { 957 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); 958 bool is_stable_ary = FoldStableValues && 959 (tp != NULL) && (tp->isa_aryptr() != NULL) && 960 tp->isa_aryptr()->is_stable(); 961 962 return (eliminate_boxing && non_volatile) || is_stable_ary; 963 } 964 965 return false; 966 } 967 968 //---------------------------can_see_stored_value------------------------------ 969 // This routine exists to make sure this set of tests is done the same 970 // everywhere. We need to make a coordinated change: first LoadNode::Ideal 971 // will change the graph shape in a way which makes memory alive twice at the 972 // same time (uses the Oracle model of aliasing), then some 973 // LoadXNode::Identity will fold things back to the equivalence-class model 974 // of aliasing. 975 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { 976 Node* ld_adr = in(MemNode::Address); 977 intptr_t ld_off = 0; 978 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); 979 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); 980 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; 981 // This is more general than load from boxing objects. 982 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { 983 uint alias_idx = atp->index(); 984 bool final = !atp->is_rewritable(); 985 Node* result = NULL; 986 Node* current = st; 987 // Skip through chains of MemBarNodes checking the MergeMems for 988 // new states for the slice of this load. Stop once any other 989 // kind of node is encountered. Loads from final memory can skip 990 // through any kind of MemBar but normal loads shouldn't skip 991 // through MemBarAcquire since the could allow them to move out of 992 // a synchronized region. 993 while (current->is_Proj()) { 994 int opc = current->in(0)->Opcode(); 995 if ((final && (opc == Op_MemBarAcquire || 996 opc == Op_MemBarAcquireLock || 997 opc == Op_LoadFence)) || 998 opc == Op_MemBarRelease || 999 opc == Op_StoreFence || 1000 opc == Op_MemBarReleaseLock || 1001 opc == Op_MemBarCPUOrder) { 1002 Node* mem = current->in(0)->in(TypeFunc::Memory); 1003 if (mem->is_MergeMem()) { 1004 MergeMemNode* merge = mem->as_MergeMem(); 1005 Node* new_st = merge->memory_at(alias_idx); 1006 if (new_st == merge->base_memory()) { 1007 // Keep searching 1008 current = new_st; 1009 continue; 1010 } 1011 // Save the new memory state for the slice and fall through 1012 // to exit. 1013 result = new_st; 1014 } 1015 } 1016 break; 1017 } 1018 if (result != NULL) { 1019 st = result; 1020 } 1021 } 1022 1023 // Loop around twice in the case Load -> Initialize -> Store. 1024 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) 1025 for (int trip = 0; trip <= 1; trip++) { 1026 1027 if (st->is_Store()) { 1028 Node* st_adr = st->in(MemNode::Address); 1029 if (!phase->eqv(st_adr, ld_adr)) { 1030 // Try harder before giving up... Match raw and non-raw pointers. 1031 intptr_t st_off = 0; 1032 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); 1033 if (alloc == NULL) return NULL; 1034 if (alloc != ld_alloc) return NULL; 1035 if (ld_off != st_off) return NULL; 1036 // At this point we have proven something like this setup: 1037 // A = Allocate(...) 1038 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) 1039 // S = StoreQ(, AddP(, A.Parm , #Off), V) 1040 // (Actually, we haven't yet proven the Q's are the same.) 1041 // In other words, we are loading from a casted version of 1042 // the same pointer-and-offset that we stored to. 1043 // Thus, we are able to replace L by V. 1044 } 1045 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. 1046 if (store_Opcode() != st->Opcode()) 1047 return NULL; 1048 return st->in(MemNode::ValueIn); 1049 } 1050 1051 // A load from a freshly-created object always returns zero. 1052 // (This can happen after LoadNode::Ideal resets the load's memory input 1053 // to find_captured_store, which returned InitializeNode::zero_memory.) 1054 if (st->is_Proj() && st->in(0)->is_Allocate() && 1055 (st->in(0) == ld_alloc) && 1056 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { 1057 // return a zero value for the load's basic type 1058 // (This is one of the few places where a generic PhaseTransform 1059 // can create new nodes. Think of it as lazily manifesting 1060 // virtually pre-existing constants.) 1061 return phase->zerocon(memory_type()); 1062 } 1063 1064 // A load from an initialization barrier can match a captured store. 1065 if (st->is_Proj() && st->in(0)->is_Initialize()) { 1066 InitializeNode* init = st->in(0)->as_Initialize(); 1067 AllocateNode* alloc = init->allocation(); 1068 if ((alloc != NULL) && (alloc == ld_alloc)) { 1069 // examine a captured store value 1070 st = init->find_captured_store(ld_off, memory_size(), phase); 1071 if (st != NULL) 1072 continue; // take one more trip around 1073 } 1074 } 1075 1076 // Load boxed value from result of valueOf() call is input parameter. 1077 if (this->is_Load() && ld_adr->is_AddP() && 1078 (tp != NULL) && tp->is_ptr_to_boxed_value()) { 1079 intptr_t ignore = 0; 1080 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); 1081 if (base != NULL && base->is_Proj() && 1082 base->as_Proj()->_con == TypeFunc::Parms && 1083 base->in(0)->is_CallStaticJava() && 1084 base->in(0)->as_CallStaticJava()->is_boxing_method()) { 1085 return base->in(0)->in(TypeFunc::Parms); 1086 } 1087 } 1088 1089 break; 1090 } 1091 1092 return NULL; 1093 } 1094 1095 //----------------------is_instance_field_load_with_local_phi------------------ 1096 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { 1097 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && 1098 in(Address)->is_AddP() ) { 1099 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); 1100 // Only instances and boxed values. 1101 if( t_oop != NULL && 1102 (t_oop->is_ptr_to_boxed_value() || 1103 t_oop->is_known_instance_field()) && 1104 t_oop->offset() != Type::OffsetBot && 1105 t_oop->offset() != Type::OffsetTop) { 1106 return true; 1107 } 1108 } 1109 return false; 1110 } 1111 1112 //------------------------------Identity--------------------------------------- 1113 // Loads are identity if previous store is to same address 1114 Node *LoadNode::Identity( PhaseTransform *phase ) { 1115 // If the previous store-maker is the right kind of Store, and the store is 1116 // to the same address, then we are equal to the value stored. 1117 Node* mem = in(Memory); 1118 Node* value = can_see_stored_value(mem, phase); 1119 if( value ) { 1120 // byte, short & char stores truncate naturally. 1121 // A load has to load the truncated value which requires 1122 // some sort of masking operation and that requires an 1123 // Ideal call instead of an Identity call. 1124 if (memory_size() < BytesPerInt) { 1125 // If the input to the store does not fit with the load's result type, 1126 // it must be truncated via an Ideal call. 1127 if (!phase->type(value)->higher_equal(phase->type(this))) 1128 return this; 1129 } 1130 // (This works even when value is a Con, but LoadNode::Value 1131 // usually runs first, producing the singleton type of the Con.) 1132 return value; 1133 } 1134 1135 // Search for an existing data phi which was generated before for the same 1136 // instance's field to avoid infinite generation of phis in a loop. 1137 Node *region = mem->in(0); 1138 if (is_instance_field_load_with_local_phi(region)) { 1139 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); 1140 int this_index = phase->C->get_alias_index(addr_t); 1141 int this_offset = addr_t->offset(); 1142 int this_iid = addr_t->instance_id(); 1143 if (!addr_t->is_known_instance() && 1144 addr_t->is_ptr_to_boxed_value()) { 1145 // Use _idx of address base (could be Phi node) for boxed values. 1146 intptr_t ignore = 0; 1147 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1148 this_iid = base->_idx; 1149 } 1150 const Type* this_type = bottom_type(); 1151 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1152 Node* phi = region->fast_out(i); 1153 if (phi->is_Phi() && phi != mem && 1154 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { 1155 return phi; 1156 } 1157 } 1158 } 1159 1160 return this; 1161 } 1162 1163 // We're loading from an object which has autobox behaviour. 1164 // If this object is result of a valueOf call we'll have a phi 1165 // merging a newly allocated object and a load from the cache. 1166 // We want to replace this load with the original incoming 1167 // argument to the valueOf call. 1168 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { 1169 assert(phase->C->eliminate_boxing(), "sanity"); 1170 intptr_t ignore = 0; 1171 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); 1172 if ((base == NULL) || base->is_Phi()) { 1173 // Push the loads from the phi that comes from valueOf up 1174 // through it to allow elimination of the loads and the recovery 1175 // of the original value. It is done in split_through_phi(). 1176 return NULL; 1177 } else if (base->is_Load() || 1178 base->is_DecodeN() && base->in(1)->is_Load()) { 1179 // Eliminate the load of boxed value for integer types from the cache 1180 // array by deriving the value from the index into the array. 1181 // Capture the offset of the load and then reverse the computation. 1182 1183 // Get LoadN node which loads a boxing object from 'cache' array. 1184 if (base->is_DecodeN()) { 1185 base = base->in(1); 1186 } 1187 if (!base->in(Address)->is_AddP()) { 1188 return NULL; // Complex address 1189 } 1190 AddPNode* address = base->in(Address)->as_AddP(); 1191 Node* cache_base = address->in(AddPNode::Base); 1192 if ((cache_base != NULL) && cache_base->is_DecodeN()) { 1193 // Get ConP node which is static 'cache' field. 1194 cache_base = cache_base->in(1); 1195 } 1196 if ((cache_base != NULL) && cache_base->is_Con()) { 1197 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); 1198 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1199 Node* elements[4]; 1200 int shift = exact_log2(type2aelembytes(T_OBJECT)); 1201 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); 1202 if ((count > 0) && elements[0]->is_Con() && 1203 ((count == 1) || 1204 (count == 2) && elements[1]->Opcode() == Op_LShiftX && 1205 elements[1]->in(2) == phase->intcon(shift))) { 1206 ciObjArray* array = base_type->const_oop()->as_obj_array(); 1207 // Fetch the box object cache[0] at the base of the array and get its value 1208 ciInstance* box = array->obj_at(0)->as_instance(); 1209 ciInstanceKlass* ik = box->klass()->as_instance_klass(); 1210 assert(ik->is_box_klass(), "sanity"); 1211 assert(ik->nof_nonstatic_fields() == 1, "change following code"); 1212 if (ik->nof_nonstatic_fields() == 1) { 1213 // This should be true nonstatic_field_at requires calling 1214 // nof_nonstatic_fields so check it anyway 1215 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); 1216 BasicType bt = c.basic_type(); 1217 // Only integer types have boxing cache. 1218 assert(bt == T_BOOLEAN || bt == T_CHAR || 1219 bt == T_BYTE || bt == T_SHORT || 1220 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); 1221 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); 1222 if (cache_low != (int)cache_low) { 1223 return NULL; // should not happen since cache is array indexed by value 1224 } 1225 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); 1226 if (offset != (int)offset) { 1227 return NULL; // should not happen since cache is array indexed by value 1228 } 1229 // Add up all the offsets making of the address of the load 1230 Node* result = elements[0]; 1231 for (int i = 1; i < count; i++) { 1232 result = phase->transform(new AddXNode(result, elements[i])); 1233 } 1234 // Remove the constant offset from the address and then 1235 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset))); 1236 // remove the scaling of the offset to recover the original index. 1237 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { 1238 // Peel the shift off directly but wrap it in a dummy node 1239 // since Ideal can't return existing nodes 1240 result = new RShiftXNode(result->in(1), phase->intcon(0)); 1241 } else if (result->is_Add() && result->in(2)->is_Con() && 1242 result->in(1)->Opcode() == Op_LShiftX && 1243 result->in(1)->in(2) == phase->intcon(shift)) { 1244 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) 1245 // but for boxing cache access we know that X<<Z will not overflow 1246 // (there is range check) so we do this optimizatrion by hand here. 1247 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift)); 1248 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con)); 1249 } else { 1250 result = new RShiftXNode(result, phase->intcon(shift)); 1251 } 1252 #ifdef _LP64 1253 if (bt != T_LONG) { 1254 result = new ConvL2INode(phase->transform(result)); 1255 } 1256 #else 1257 if (bt == T_LONG) { 1258 result = new ConvI2LNode(phase->transform(result)); 1259 } 1260 #endif 1261 return result; 1262 } 1263 } 1264 } 1265 } 1266 } 1267 return NULL; 1268 } 1269 1270 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { 1271 Node* region = phi->in(0); 1272 if (region == NULL) { 1273 return false; // Wait stable graph 1274 } 1275 uint cnt = phi->req(); 1276 for (uint i = 1; i < cnt; i++) { 1277 Node* rc = region->in(i); 1278 if (rc == NULL || phase->type(rc) == Type::TOP) 1279 return false; // Wait stable graph 1280 Node* in = phi->in(i); 1281 if (in == NULL || phase->type(in) == Type::TOP) 1282 return false; // Wait stable graph 1283 } 1284 return true; 1285 } 1286 //------------------------------split_through_phi------------------------------ 1287 // Split instance or boxed field load through Phi. 1288 Node *LoadNode::split_through_phi(PhaseGVN *phase) { 1289 Node* mem = in(Memory); 1290 Node* address = in(Address); 1291 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); 1292 1293 assert((t_oop != NULL) && 1294 (t_oop->is_known_instance_field() || 1295 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); 1296 1297 Compile* C = phase->C; 1298 intptr_t ignore = 0; 1299 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1300 bool base_is_phi = (base != NULL) && base->is_Phi(); 1301 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && 1302 (base != NULL) && (base == address->in(AddPNode::Base)) && 1303 phase->type(base)->higher_equal(TypePtr::NOTNULL); 1304 1305 if (!((mem->is_Phi() || base_is_phi) && 1306 (load_boxed_values || t_oop->is_known_instance_field()))) { 1307 return NULL; // memory is not Phi 1308 } 1309 1310 if (mem->is_Phi()) { 1311 if (!stable_phi(mem->as_Phi(), phase)) { 1312 return NULL; // Wait stable graph 1313 } 1314 uint cnt = mem->req(); 1315 // Check for loop invariant memory. 1316 if (cnt == 3) { 1317 for (uint i = 1; i < cnt; i++) { 1318 Node* in = mem->in(i); 1319 Node* m = optimize_memory_chain(in, t_oop, this, phase); 1320 if (m == mem) { 1321 set_req(Memory, mem->in(cnt - i)); 1322 return this; // made change 1323 } 1324 } 1325 } 1326 } 1327 if (base_is_phi) { 1328 if (!stable_phi(base->as_Phi(), phase)) { 1329 return NULL; // Wait stable graph 1330 } 1331 uint cnt = base->req(); 1332 // Check for loop invariant memory. 1333 if (cnt == 3) { 1334 for (uint i = 1; i < cnt; i++) { 1335 if (base->in(i) == base) { 1336 return NULL; // Wait stable graph 1337 } 1338 } 1339 } 1340 } 1341 1342 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); 1343 1344 // Split through Phi (see original code in loopopts.cpp). 1345 assert(C->have_alias_type(t_oop), "instance should have alias type"); 1346 1347 // Do nothing here if Identity will find a value 1348 // (to avoid infinite chain of value phis generation). 1349 if (!phase->eqv(this, this->Identity(phase))) 1350 return NULL; 1351 1352 // Select Region to split through. 1353 Node* region; 1354 if (!base_is_phi) { 1355 assert(mem->is_Phi(), "sanity"); 1356 region = mem->in(0); 1357 // Skip if the region dominates some control edge of the address. 1358 if (!MemNode::all_controls_dominate(address, region)) 1359 return NULL; 1360 } else if (!mem->is_Phi()) { 1361 assert(base_is_phi, "sanity"); 1362 region = base->in(0); 1363 // Skip if the region dominates some control edge of the memory. 1364 if (!MemNode::all_controls_dominate(mem, region)) 1365 return NULL; 1366 } else if (base->in(0) != mem->in(0)) { 1367 assert(base_is_phi && mem->is_Phi(), "sanity"); 1368 if (MemNode::all_controls_dominate(mem, base->in(0))) { 1369 region = base->in(0); 1370 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { 1371 region = mem->in(0); 1372 } else { 1373 return NULL; // complex graph 1374 } 1375 } else { 1376 assert(base->in(0) == mem->in(0), "sanity"); 1377 region = mem->in(0); 1378 } 1379 1380 const Type* this_type = this->bottom_type(); 1381 int this_index = C->get_alias_index(t_oop); 1382 int this_offset = t_oop->offset(); 1383 int this_iid = t_oop->instance_id(); 1384 if (!t_oop->is_known_instance() && load_boxed_values) { 1385 // Use _idx of address base for boxed values. 1386 this_iid = base->_idx; 1387 } 1388 PhaseIterGVN* igvn = phase->is_IterGVN(); 1389 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); 1390 for (uint i = 1; i < region->req(); i++) { 1391 Node* x; 1392 Node* the_clone = NULL; 1393 if (region->in(i) == C->top()) { 1394 x = C->top(); // Dead path? Use a dead data op 1395 } else { 1396 x = this->clone(); // Else clone up the data op 1397 the_clone = x; // Remember for possible deletion. 1398 // Alter data node to use pre-phi inputs 1399 if (this->in(0) == region) { 1400 x->set_req(0, region->in(i)); 1401 } else { 1402 x->set_req(0, NULL); 1403 } 1404 if (mem->is_Phi() && (mem->in(0) == region)) { 1405 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. 1406 } 1407 if (address->is_Phi() && address->in(0) == region) { 1408 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone 1409 } 1410 if (base_is_phi && (base->in(0) == region)) { 1411 Node* base_x = base->in(i); // Clone address for loads from boxed objects. 1412 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset))); 1413 x->set_req(Address, adr_x); 1414 } 1415 } 1416 // Check for a 'win' on some paths 1417 const Type *t = x->Value(igvn); 1418 1419 bool singleton = t->singleton(); 1420 1421 // See comments in PhaseIdealLoop::split_thru_phi(). 1422 if (singleton && t == Type::TOP) { 1423 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 1424 } 1425 1426 if (singleton) { 1427 x = igvn->makecon(t); 1428 } else { 1429 // We now call Identity to try to simplify the cloned node. 1430 // Note that some Identity methods call phase->type(this). 1431 // Make sure that the type array is big enough for 1432 // our new node, even though we may throw the node away. 1433 // (This tweaking with igvn only works because x is a new node.) 1434 igvn->set_type(x, t); 1435 // If x is a TypeNode, capture any more-precise type permanently into Node 1436 // otherwise it will be not updated during igvn->transform since 1437 // igvn->type(x) is set to x->Value() already. 1438 x->raise_bottom_type(t); 1439 Node *y = x->Identity(igvn); 1440 if (y != x) { 1441 x = y; 1442 } else { 1443 y = igvn->hash_find_insert(x); 1444 if (y) { 1445 x = y; 1446 } else { 1447 // Else x is a new node we are keeping 1448 // We do not need register_new_node_with_optimizer 1449 // because set_type has already been called. 1450 igvn->_worklist.push(x); 1451 } 1452 } 1453 } 1454 if (x != the_clone && the_clone != NULL) { 1455 igvn->remove_dead_node(the_clone); 1456 } 1457 phi->set_req(i, x); 1458 } 1459 // Record Phi 1460 igvn->register_new_node_with_optimizer(phi); 1461 return phi; 1462 } 1463 1464 //------------------------------Ideal------------------------------------------ 1465 // If the load is from Field memory and the pointer is non-null, we can 1466 // zero out the control input. 1467 // If the offset is constant and the base is an object allocation, 1468 // try to hook me up to the exact initializing store. 1469 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1470 Node* p = MemNode::Ideal_common(phase, can_reshape); 1471 if (p) return (p == NodeSentinel) ? NULL : p; 1472 1473 Node* ctrl = in(MemNode::Control); 1474 Node* address = in(MemNode::Address); 1475 1476 // Skip up past a SafePoint control. Cannot do this for Stores because 1477 // pointer stores & cardmarks must stay on the same side of a SafePoint. 1478 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && 1479 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { 1480 ctrl = ctrl->in(0); 1481 set_req(MemNode::Control,ctrl); 1482 } 1483 1484 intptr_t ignore = 0; 1485 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); 1486 if (base != NULL 1487 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { 1488 // Check for useless control edge in some common special cases 1489 if (in(MemNode::Control) != NULL 1490 && phase->type(base)->higher_equal(TypePtr::NOTNULL) 1491 && all_controls_dominate(base, phase->C->start())) { 1492 // A method-invariant, non-null address (constant or 'this' argument). 1493 set_req(MemNode::Control, NULL); 1494 } 1495 } 1496 1497 Node* mem = in(MemNode::Memory); 1498 const TypePtr *addr_t = phase->type(address)->isa_ptr(); 1499 1500 if (can_reshape && (addr_t != NULL)) { 1501 // try to optimize our memory input 1502 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); 1503 if (opt_mem != mem) { 1504 set_req(MemNode::Memory, opt_mem); 1505 if (phase->type( opt_mem ) == Type::TOP) return NULL; 1506 return this; 1507 } 1508 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); 1509 if ((t_oop != NULL) && 1510 (t_oop->is_known_instance_field() || 1511 t_oop->is_ptr_to_boxed_value())) { 1512 PhaseIterGVN *igvn = phase->is_IterGVN(); 1513 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { 1514 // Delay this transformation until memory Phi is processed. 1515 phase->is_IterGVN()->_worklist.push(this); 1516 return NULL; 1517 } 1518 // Split instance field load through Phi. 1519 Node* result = split_through_phi(phase); 1520 if (result != NULL) return result; 1521 1522 if (t_oop->is_ptr_to_boxed_value()) { 1523 Node* result = eliminate_autobox(phase); 1524 if (result != NULL) return result; 1525 } 1526 } 1527 } 1528 1529 // Check for prior store with a different base or offset; make Load 1530 // independent. Skip through any number of them. Bail out if the stores 1531 // are in an endless dead cycle and report no progress. This is a key 1532 // transform for Reflection. However, if after skipping through the Stores 1533 // we can't then fold up against a prior store do NOT do the transform as 1534 // this amounts to using the 'Oracle' model of aliasing. It leaves the same 1535 // array memory alive twice: once for the hoisted Load and again after the 1536 // bypassed Store. This situation only works if EVERYBODY who does 1537 // anti-dependence work knows how to bypass. I.e. we need all 1538 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is 1539 // the alias index stuff. So instead, peek through Stores and IFF we can 1540 // fold up, do so. 1541 Node* prev_mem = find_previous_store(phase); 1542 // Steps (a), (b): Walk past independent stores to find an exact match. 1543 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { 1544 // (c) See if we can fold up on the spot, but don't fold up here. 1545 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or 1546 // just return a prior value, which is done by Identity calls. 1547 if (can_see_stored_value(prev_mem, phase)) { 1548 // Make ready for step (d): 1549 set_req(MemNode::Memory, prev_mem); 1550 return this; 1551 } 1552 } 1553 1554 return NULL; // No further progress 1555 } 1556 1557 // Helper to recognize certain Klass fields which are invariant across 1558 // some group of array types (e.g., int[] or all T[] where T < Object). 1559 const Type* 1560 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, 1561 ciKlass* klass) const { 1562 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { 1563 // The field is Klass::_modifier_flags. Return its (constant) value. 1564 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) 1565 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); 1566 return TypeInt::make(klass->modifier_flags()); 1567 } 1568 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { 1569 // The field is Klass::_access_flags. Return its (constant) value. 1570 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) 1571 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); 1572 return TypeInt::make(klass->access_flags()); 1573 } 1574 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { 1575 // The field is Klass::_layout_helper. Return its constant value if known. 1576 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1577 return TypeInt::make(klass->layout_helper()); 1578 } 1579 1580 // No match. 1581 return NULL; 1582 } 1583 1584 // Try to constant-fold a stable array element. 1585 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { 1586 assert(ary->const_oop(), "array should be constant"); 1587 assert(ary->is_stable(), "array should be stable"); 1588 1589 // Decode the results of GraphKit::array_element_address. 1590 ciArray* aobj = ary->const_oop()->as_array(); 1591 ciConstant con = aobj->element_value_by_offset(off); 1592 1593 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { 1594 const Type* con_type = Type::make_from_constant(con); 1595 if (con_type != NULL) { 1596 if (con_type->isa_aryptr()) { 1597 // Join with the array element type, in case it is also stable. 1598 int dim = ary->stable_dimension(); 1599 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); 1600 } 1601 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { 1602 con_type = con_type->make_narrowoop(); 1603 } 1604 #ifndef PRODUCT 1605 if (TraceIterativeGVN) { 1606 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); 1607 con_type->dump(); tty->cr(); 1608 } 1609 #endif //PRODUCT 1610 return con_type; 1611 } 1612 } 1613 return NULL; 1614 } 1615 1616 //------------------------------Value----------------------------------------- 1617 const Type *LoadNode::Value( PhaseTransform *phase ) const { 1618 // Either input is TOP ==> the result is TOP 1619 Node* mem = in(MemNode::Memory); 1620 const Type *t1 = phase->type(mem); 1621 if (t1 == Type::TOP) return Type::TOP; 1622 Node* adr = in(MemNode::Address); 1623 const TypePtr* tp = phase->type(adr)->isa_ptr(); 1624 if (tp == NULL || tp->empty()) return Type::TOP; 1625 int off = tp->offset(); 1626 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); 1627 Compile* C = phase->C; 1628 1629 // Try to guess loaded type from pointer type 1630 if (tp->isa_aryptr()) { 1631 const TypeAryPtr* ary = tp->is_aryptr(); 1632 const Type* t = ary->elem(); 1633 1634 // Determine whether the reference is beyond the header or not, by comparing 1635 // the offset against the offset of the start of the array's data. 1636 // Different array types begin at slightly different offsets (12 vs. 16). 1637 // We choose T_BYTE as an example base type that is least restrictive 1638 // as to alignment, which will therefore produce the smallest 1639 // possible base offset. 1640 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1641 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); 1642 1643 // Try to constant-fold a stable array element. 1644 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { 1645 // Make sure the reference is not into the header and the offset is constant 1646 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { 1647 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); 1648 if (con_type != NULL) { 1649 return con_type; 1650 } 1651 } 1652 } 1653 1654 // Don't do this for integer types. There is only potential profit if 1655 // the element type t is lower than _type; that is, for int types, if _type is 1656 // more restrictive than t. This only happens here if one is short and the other 1657 // char (both 16 bits), and in those cases we've made an intentional decision 1658 // to use one kind of load over the other. See AndINode::Ideal and 4965907. 1659 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. 1660 // 1661 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) 1662 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier 1663 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been 1664 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, 1665 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. 1666 // In fact, that could have been the original type of p1, and p1 could have 1667 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the 1668 // expression (LShiftL quux 3) independently optimized to the constant 8. 1669 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) 1670 && (_type->isa_vect() == NULL) 1671 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { 1672 // t might actually be lower than _type, if _type is a unique 1673 // concrete subclass of abstract class t. 1674 if (off_beyond_header) { // is the offset beyond the header? 1675 const Type* jt = t->join_speculative(_type); 1676 // In any case, do not allow the join, per se, to empty out the type. 1677 if (jt->empty() && !t->empty()) { 1678 // This can happen if a interface-typed array narrows to a class type. 1679 jt = _type; 1680 } 1681 #ifdef ASSERT 1682 if (phase->C->eliminate_boxing() && adr->is_AddP()) { 1683 // The pointers in the autobox arrays are always non-null 1684 Node* base = adr->in(AddPNode::Base); 1685 if ((base != NULL) && base->is_DecodeN()) { 1686 // Get LoadN node which loads IntegerCache.cache field 1687 base = base->in(1); 1688 } 1689 if ((base != NULL) && base->is_Con()) { 1690 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); 1691 if ((base_type != NULL) && base_type->is_autobox_cache()) { 1692 // It could be narrow oop 1693 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); 1694 } 1695 } 1696 } 1697 #endif 1698 return jt; 1699 } 1700 } 1701 } else if (tp->base() == Type::InstPtr) { 1702 ciEnv* env = C->env(); 1703 const TypeInstPtr* tinst = tp->is_instptr(); 1704 ciKlass* klass = tinst->klass(); 1705 assert( off != Type::OffsetBot || 1706 // arrays can be cast to Objects 1707 tp->is_oopptr()->klass()->is_java_lang_Object() || 1708 // unsafe field access may not have a constant offset 1709 C->has_unsafe_access(), 1710 "Field accesses must be precise" ); 1711 // For oop loads, we expect the _type to be precise 1712 if (klass == env->String_klass() && 1713 adr->is_AddP() && off != Type::OffsetBot) { 1714 // For constant Strings treat the final fields as compile time constants. 1715 Node* base = adr->in(AddPNode::Base); 1716 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); 1717 if (t != NULL && t->singleton()) { 1718 ciField* field = env->String_klass()->get_field_by_offset(off, false); 1719 if (field != NULL && field->is_final()) { 1720 ciObject* string = t->const_oop(); 1721 ciConstant constant = string->as_instance()->field_value(field); 1722 if (constant.basic_type() == T_INT) { 1723 return TypeInt::make(constant.as_int()); 1724 } else if (constant.basic_type() == T_ARRAY) { 1725 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1726 return TypeNarrowOop::make_from_constant(constant.as_object(), true); 1727 } else { 1728 return TypeOopPtr::make_from_constant(constant.as_object(), true); 1729 } 1730 } 1731 } 1732 } 1733 } 1734 // Optimizations for constant objects 1735 ciObject* const_oop = tinst->const_oop(); 1736 if (const_oop != NULL) { 1737 // For constant Boxed value treat the target field as a compile time constant. 1738 if (tinst->is_ptr_to_boxed_value()) { 1739 return tinst->get_const_boxed_value(); 1740 } else 1741 // For constant CallSites treat the target field as a compile time constant. 1742 if (const_oop->is_call_site()) { 1743 ciCallSite* call_site = const_oop->as_call_site(); 1744 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); 1745 if (field != NULL && field->is_call_site_target()) { 1746 ciMethodHandle* target = call_site->get_target(); 1747 if (target != NULL) { // just in case 1748 ciConstant constant(T_OBJECT, target); 1749 const Type* t; 1750 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 1751 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); 1752 } else { 1753 t = TypeOopPtr::make_from_constant(constant.as_object(), true); 1754 } 1755 // Add a dependence for invalidation of the optimization. 1756 if (!call_site->is_constant_call_site()) { 1757 C->dependencies()->assert_call_site_target_value(call_site, target); 1758 } 1759 return t; 1760 } 1761 } 1762 } 1763 } 1764 } else if (tp->base() == Type::KlassPtr) { 1765 assert( off != Type::OffsetBot || 1766 // arrays can be cast to Objects 1767 tp->is_klassptr()->klass()->is_java_lang_Object() || 1768 // also allow array-loading from the primary supertype 1769 // array during subtype checks 1770 Opcode() == Op_LoadKlass, 1771 "Field accesses must be precise" ); 1772 // For klass/static loads, we expect the _type to be precise 1773 } 1774 1775 const TypeKlassPtr *tkls = tp->isa_klassptr(); 1776 if (tkls != NULL && !StressReflectiveCode) { 1777 ciKlass* klass = tkls->klass(); 1778 if (klass->is_loaded() && tkls->klass_is_exact()) { 1779 // We are loading a field from a Klass metaobject whose identity 1780 // is known at compile time (the type is "exact" or "precise"). 1781 // Check for fields we know are maintained as constants by the VM. 1782 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { 1783 // The field is Klass::_super_check_offset. Return its (constant) value. 1784 // (Folds up type checking code.) 1785 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); 1786 return TypeInt::make(klass->super_check_offset()); 1787 } 1788 // Compute index into primary_supers array 1789 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1790 // Check for overflowing; use unsigned compare to handle the negative case. 1791 if( depth < ciKlass::primary_super_limit() ) { 1792 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1793 // (Folds up type checking code.) 1794 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1795 ciKlass *ss = klass->super_of_depth(depth); 1796 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1797 } 1798 const Type* aift = load_array_final_field(tkls, klass); 1799 if (aift != NULL) return aift; 1800 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) 1801 && klass->is_array_klass()) { 1802 // The field is ArrayKlass::_component_mirror. Return its (constant) value. 1803 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) 1804 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); 1805 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); 1806 } 1807 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { 1808 // The field is Klass::_java_mirror. Return its (constant) value. 1809 // (Folds up the 2nd indirection in anObjConstant.getClass().) 1810 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); 1811 return TypeInstPtr::make(klass->java_mirror()); 1812 } 1813 } 1814 1815 // We can still check if we are loading from the primary_supers array at a 1816 // shallow enough depth. Even though the klass is not exact, entries less 1817 // than or equal to its super depth are correct. 1818 if (klass->is_loaded() ) { 1819 ciType *inner = klass; 1820 while( inner->is_obj_array_klass() ) 1821 inner = inner->as_obj_array_klass()->base_element_type(); 1822 if( inner->is_instance_klass() && 1823 !inner->as_instance_klass()->flags().is_interface() ) { 1824 // Compute index into primary_supers array 1825 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); 1826 // Check for overflowing; use unsigned compare to handle the negative case. 1827 if( depth < ciKlass::primary_super_limit() && 1828 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case 1829 // The field is an element of Klass::_primary_supers. Return its (constant) value. 1830 // (Folds up type checking code.) 1831 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); 1832 ciKlass *ss = klass->super_of_depth(depth); 1833 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; 1834 } 1835 } 1836 } 1837 1838 // If the type is enough to determine that the thing is not an array, 1839 // we can give the layout_helper a positive interval type. 1840 // This will help short-circuit some reflective code. 1841 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) 1842 && !klass->is_array_klass() // not directly typed as an array 1843 && !klass->is_interface() // specifically not Serializable & Cloneable 1844 && !klass->is_java_lang_Object() // not the supertype of all T[] 1845 ) { 1846 // Note: When interfaces are reliable, we can narrow the interface 1847 // test to (klass != Serializable && klass != Cloneable). 1848 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); 1849 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); 1850 // The key property of this type is that it folds up tests 1851 // for array-ness, since it proves that the layout_helper is positive. 1852 // Thus, a generic value like the basic object layout helper works fine. 1853 return TypeInt::make(min_size, max_jint, Type::WidenMin); 1854 } 1855 } 1856 1857 // If we are loading from a freshly-allocated object, produce a zero, 1858 // if the load is provably beyond the header of the object. 1859 // (Also allow a variable load from a fresh array to produce zero.) 1860 const TypeOopPtr *tinst = tp->isa_oopptr(); 1861 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); 1862 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); 1863 if (ReduceFieldZeroing || is_instance || is_boxed_value) { 1864 Node* value = can_see_stored_value(mem,phase); 1865 if (value != NULL && value->is_Con()) { 1866 assert(value->bottom_type()->higher_equal(_type),"sanity"); 1867 return value->bottom_type(); 1868 } 1869 } 1870 1871 if (is_instance) { 1872 // If we have an instance type and our memory input is the 1873 // programs's initial memory state, there is no matching store, 1874 // so just return a zero of the appropriate type 1875 Node *mem = in(MemNode::Memory); 1876 if (mem->is_Parm() && mem->in(0)->is_Start()) { 1877 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); 1878 return Type::get_zero_type(_type->basic_type()); 1879 } 1880 } 1881 return _type; 1882 } 1883 1884 //------------------------------match_edge------------------------------------- 1885 // Do we Match on this edge index or not? Match only the address. 1886 uint LoadNode::match_edge(uint idx) const { 1887 return idx == MemNode::Address; 1888 } 1889 1890 //--------------------------LoadBNode::Ideal-------------------------------------- 1891 // 1892 // If the previous store is to the same address as this load, 1893 // and the value stored was larger than a byte, replace this load 1894 // with the value stored truncated to a byte. If no truncation is 1895 // needed, the replacement is done in LoadNode::Identity(). 1896 // 1897 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1898 Node* mem = in(MemNode::Memory); 1899 Node* value = can_see_stored_value(mem,phase); 1900 if( value && !phase->type(value)->higher_equal( _type ) ) { 1901 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) ); 1902 return new RShiftINode(result, phase->intcon(24)); 1903 } 1904 // Identity call will handle the case where truncation is not needed. 1905 return LoadNode::Ideal(phase, can_reshape); 1906 } 1907 1908 const Type* LoadBNode::Value(PhaseTransform *phase) const { 1909 Node* mem = in(MemNode::Memory); 1910 Node* value = can_see_stored_value(mem,phase); 1911 if (value != NULL && value->is_Con() && 1912 !value->bottom_type()->higher_equal(_type)) { 1913 // If the input to the store does not fit with the load's result type, 1914 // it must be truncated. We can't delay until Ideal call since 1915 // a singleton Value is needed for split_thru_phi optimization. 1916 int con = value->get_int(); 1917 return TypeInt::make((con << 24) >> 24); 1918 } 1919 return LoadNode::Value(phase); 1920 } 1921 1922 //--------------------------LoadUBNode::Ideal------------------------------------- 1923 // 1924 // If the previous store is to the same address as this load, 1925 // and the value stored was larger than a byte, replace this load 1926 // with the value stored truncated to a byte. If no truncation is 1927 // needed, the replacement is done in LoadNode::Identity(). 1928 // 1929 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1930 Node* mem = in(MemNode::Memory); 1931 Node* value = can_see_stored_value(mem, phase); 1932 if (value && !phase->type(value)->higher_equal(_type)) 1933 return new AndINode(value, phase->intcon(0xFF)); 1934 // Identity call will handle the case where truncation is not needed. 1935 return LoadNode::Ideal(phase, can_reshape); 1936 } 1937 1938 const Type* LoadUBNode::Value(PhaseTransform *phase) const { 1939 Node* mem = in(MemNode::Memory); 1940 Node* value = can_see_stored_value(mem,phase); 1941 if (value != NULL && value->is_Con() && 1942 !value->bottom_type()->higher_equal(_type)) { 1943 // If the input to the store does not fit with the load's result type, 1944 // it must be truncated. We can't delay until Ideal call since 1945 // a singleton Value is needed for split_thru_phi optimization. 1946 int con = value->get_int(); 1947 return TypeInt::make(con & 0xFF); 1948 } 1949 return LoadNode::Value(phase); 1950 } 1951 1952 //--------------------------LoadUSNode::Ideal------------------------------------- 1953 // 1954 // If the previous store is to the same address as this load, 1955 // and the value stored was larger than a char, replace this load 1956 // with the value stored truncated to a char. If no truncation is 1957 // needed, the replacement is done in LoadNode::Identity(). 1958 // 1959 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1960 Node* mem = in(MemNode::Memory); 1961 Node* value = can_see_stored_value(mem,phase); 1962 if( value && !phase->type(value)->higher_equal( _type ) ) 1963 return new AndINode(value,phase->intcon(0xFFFF)); 1964 // Identity call will handle the case where truncation is not needed. 1965 return LoadNode::Ideal(phase, can_reshape); 1966 } 1967 1968 const Type* LoadUSNode::Value(PhaseTransform *phase) const { 1969 Node* mem = in(MemNode::Memory); 1970 Node* value = can_see_stored_value(mem,phase); 1971 if (value != NULL && value->is_Con() && 1972 !value->bottom_type()->higher_equal(_type)) { 1973 // If the input to the store does not fit with the load's result type, 1974 // it must be truncated. We can't delay until Ideal call since 1975 // a singleton Value is needed for split_thru_phi optimization. 1976 int con = value->get_int(); 1977 return TypeInt::make(con & 0xFFFF); 1978 } 1979 return LoadNode::Value(phase); 1980 } 1981 1982 //--------------------------LoadSNode::Ideal-------------------------------------- 1983 // 1984 // If the previous store is to the same address as this load, 1985 // and the value stored was larger than a short, replace this load 1986 // with the value stored truncated to a short. If no truncation is 1987 // needed, the replacement is done in LoadNode::Identity(). 1988 // 1989 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1990 Node* mem = in(MemNode::Memory); 1991 Node* value = can_see_stored_value(mem,phase); 1992 if( value && !phase->type(value)->higher_equal( _type ) ) { 1993 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) ); 1994 return new RShiftINode(result, phase->intcon(16)); 1995 } 1996 // Identity call will handle the case where truncation is not needed. 1997 return LoadNode::Ideal(phase, can_reshape); 1998 } 1999 2000 const Type* LoadSNode::Value(PhaseTransform *phase) const { 2001 Node* mem = in(MemNode::Memory); 2002 Node* value = can_see_stored_value(mem,phase); 2003 if (value != NULL && value->is_Con() && 2004 !value->bottom_type()->higher_equal(_type)) { 2005 // If the input to the store does not fit with the load's result type, 2006 // it must be truncated. We can't delay until Ideal call since 2007 // a singleton Value is needed for split_thru_phi optimization. 2008 int con = value->get_int(); 2009 return TypeInt::make((con << 16) >> 16); 2010 } 2011 return LoadNode::Value(phase); 2012 } 2013 2014 //============================================================================= 2015 //----------------------------LoadKlassNode::make------------------------------ 2016 // Polymorphic factory method: 2017 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { 2018 Compile* C = gvn.C; 2019 Node *ctl = NULL; 2020 // sanity check the alias category against the created node type 2021 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); 2022 assert(adr_type != NULL, "expecting TypeKlassPtr"); 2023 #ifdef _LP64 2024 if (adr_type->is_ptr_to_narrowklass()) { 2025 assert(UseCompressedClassPointers, "no compressed klasses"); 2026 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); 2027 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); 2028 } 2029 #endif 2030 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); 2031 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); 2032 } 2033 2034 //------------------------------Value------------------------------------------ 2035 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { 2036 return klass_value_common(phase); 2037 } 2038 2039 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { 2040 // Either input is TOP ==> the result is TOP 2041 const Type *t1 = phase->type( in(MemNode::Memory) ); 2042 if (t1 == Type::TOP) return Type::TOP; 2043 Node *adr = in(MemNode::Address); 2044 const Type *t2 = phase->type( adr ); 2045 if (t2 == Type::TOP) return Type::TOP; 2046 const TypePtr *tp = t2->is_ptr(); 2047 if (TypePtr::above_centerline(tp->ptr()) || 2048 tp->ptr() == TypePtr::Null) return Type::TOP; 2049 2050 // Return a more precise klass, if possible 2051 const TypeInstPtr *tinst = tp->isa_instptr(); 2052 if (tinst != NULL) { 2053 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 2054 int offset = tinst->offset(); 2055 if (ik == phase->C->env()->Class_klass() 2056 && (offset == java_lang_Class::klass_offset_in_bytes() || 2057 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2058 // We are loading a special hidden field from a Class mirror object, 2059 // the field which points to the VM's Klass metaobject. 2060 ciType* t = tinst->java_mirror_type(); 2061 // java_mirror_type returns non-null for compile-time Class constants. 2062 if (t != NULL) { 2063 // constant oop => constant klass 2064 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2065 if (t->is_void()) { 2066 // We cannot create a void array. Since void is a primitive type return null 2067 // klass. Users of this result need to do a null check on the returned klass. 2068 return TypePtr::NULL_PTR; 2069 } 2070 return TypeKlassPtr::make(ciArrayKlass::make(t)); 2071 } 2072 if (!t->is_klass()) { 2073 // a primitive Class (e.g., int.class) has NULL for a klass field 2074 return TypePtr::NULL_PTR; 2075 } 2076 // (Folds up the 1st indirection in aClassConstant.getModifiers().) 2077 return TypeKlassPtr::make(t->as_klass()); 2078 } 2079 // non-constant mirror, so we can't tell what's going on 2080 } 2081 if( !ik->is_loaded() ) 2082 return _type; // Bail out if not loaded 2083 if (offset == oopDesc::klass_offset_in_bytes()) { 2084 if (tinst->klass_is_exact()) { 2085 return TypeKlassPtr::make(ik); 2086 } 2087 // See if we can become precise: no subklasses and no interface 2088 // (Note: We need to support verified interfaces.) 2089 if (!ik->is_interface() && !ik->has_subklass()) { 2090 //assert(!UseExactTypes, "this code should be useless with exact types"); 2091 // Add a dependence; if any subclass added we need to recompile 2092 if (!ik->is_final()) { 2093 // %%% should use stronger assert_unique_concrete_subtype instead 2094 phase->C->dependencies()->assert_leaf_type(ik); 2095 } 2096 // Return precise klass 2097 return TypeKlassPtr::make(ik); 2098 } 2099 2100 // Return root of possible klass 2101 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); 2102 } 2103 } 2104 2105 // Check for loading klass from an array 2106 const TypeAryPtr *tary = tp->isa_aryptr(); 2107 if( tary != NULL ) { 2108 ciKlass *tary_klass = tary->klass(); 2109 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP 2110 && tary->offset() == oopDesc::klass_offset_in_bytes()) { 2111 if (tary->klass_is_exact()) { 2112 return TypeKlassPtr::make(tary_klass); 2113 } 2114 ciArrayKlass *ak = tary->klass()->as_array_klass(); 2115 // If the klass is an object array, we defer the question to the 2116 // array component klass. 2117 if( ak->is_obj_array_klass() ) { 2118 assert( ak->is_loaded(), "" ); 2119 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); 2120 if( base_k->is_loaded() && base_k->is_instance_klass() ) { 2121 ciInstanceKlass* ik = base_k->as_instance_klass(); 2122 // See if we can become precise: no subklasses and no interface 2123 if (!ik->is_interface() && !ik->has_subklass()) { 2124 //assert(!UseExactTypes, "this code should be useless with exact types"); 2125 // Add a dependence; if any subclass added we need to recompile 2126 if (!ik->is_final()) { 2127 phase->C->dependencies()->assert_leaf_type(ik); 2128 } 2129 // Return precise array klass 2130 return TypeKlassPtr::make(ak); 2131 } 2132 } 2133 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 2134 } else { // Found a type-array? 2135 //assert(!UseExactTypes, "this code should be useless with exact types"); 2136 assert( ak->is_type_array_klass(), "" ); 2137 return TypeKlassPtr::make(ak); // These are always precise 2138 } 2139 } 2140 } 2141 2142 // Check for loading klass from an array klass 2143 const TypeKlassPtr *tkls = tp->isa_klassptr(); 2144 if (tkls != NULL && !StressReflectiveCode) { 2145 ciKlass* klass = tkls->klass(); 2146 if( !klass->is_loaded() ) 2147 return _type; // Bail out if not loaded 2148 if( klass->is_obj_array_klass() && 2149 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { 2150 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 2151 // // Always returning precise element type is incorrect, 2152 // // e.g., element type could be object and array may contain strings 2153 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); 2154 2155 // The array's TypeKlassPtr was declared 'precise' or 'not precise' 2156 // according to the element type's subclassing. 2157 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); 2158 } 2159 if( klass->is_instance_klass() && tkls->klass_is_exact() && 2160 tkls->offset() == in_bytes(Klass::super_offset())) { 2161 ciKlass* sup = klass->as_instance_klass()->super(); 2162 // The field is Klass::_super. Return its (constant) value. 2163 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) 2164 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; 2165 } 2166 } 2167 2168 // Bailout case 2169 return LoadNode::Value(phase); 2170 } 2171 2172 //------------------------------Identity--------------------------------------- 2173 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. 2174 // Also feed through the klass in Allocate(...klass...)._klass. 2175 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { 2176 return klass_identity_common(phase); 2177 } 2178 2179 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { 2180 Node* x = LoadNode::Identity(phase); 2181 if (x != this) return x; 2182 2183 // Take apart the address into an oop and and offset. 2184 // Return 'this' if we cannot. 2185 Node* adr = in(MemNode::Address); 2186 intptr_t offset = 0; 2187 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2188 if (base == NULL) return this; 2189 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); 2190 if (toop == NULL) return this; 2191 2192 // We can fetch the klass directly through an AllocateNode. 2193 // This works even if the klass is not constant (clone or newArray). 2194 if (offset == oopDesc::klass_offset_in_bytes()) { 2195 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); 2196 if (allocated_klass != NULL) { 2197 return allocated_klass; 2198 } 2199 } 2200 2201 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. 2202 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. 2203 // See inline_native_Class_query for occurrences of these patterns. 2204 // Java Example: x.getClass().isAssignableFrom(y) 2205 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) 2206 // 2207 // This improves reflective code, often making the Class 2208 // mirror go completely dead. (Current exception: Class 2209 // mirrors may appear in debug info, but we could clean them out by 2210 // introducing a new debug info operator for Klass*.java_mirror). 2211 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() 2212 && (offset == java_lang_Class::klass_offset_in_bytes() || 2213 offset == java_lang_Class::array_klass_offset_in_bytes())) { 2214 // We are loading a special hidden field from a Class mirror, 2215 // the field which points to its Klass or ArrayKlass metaobject. 2216 if (base->is_Load()) { 2217 Node* adr2 = base->in(MemNode::Address); 2218 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); 2219 if (tkls != NULL && !tkls->empty() 2220 && (tkls->klass()->is_instance_klass() || 2221 tkls->klass()->is_array_klass()) 2222 && adr2->is_AddP() 2223 ) { 2224 int mirror_field = in_bytes(Klass::java_mirror_offset()); 2225 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { 2226 mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); 2227 } 2228 if (tkls->offset() == mirror_field) { 2229 return adr2->in(AddPNode::Base); 2230 } 2231 } 2232 } 2233 } 2234 2235 return this; 2236 } 2237 2238 2239 //------------------------------Value------------------------------------------ 2240 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { 2241 const Type *t = klass_value_common(phase); 2242 if (t == Type::TOP) 2243 return t; 2244 2245 return t->make_narrowklass(); 2246 } 2247 2248 //------------------------------Identity--------------------------------------- 2249 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. 2250 // Also feed through the klass in Allocate(...klass...)._klass. 2251 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { 2252 Node *x = klass_identity_common(phase); 2253 2254 const Type *t = phase->type( x ); 2255 if( t == Type::TOP ) return x; 2256 if( t->isa_narrowklass()) return x; 2257 assert (!t->isa_narrowoop(), "no narrow oop here"); 2258 2259 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass())); 2260 } 2261 2262 //------------------------------Value----------------------------------------- 2263 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { 2264 // Either input is TOP ==> the result is TOP 2265 const Type *t1 = phase->type( in(MemNode::Memory) ); 2266 if( t1 == Type::TOP ) return Type::TOP; 2267 Node *adr = in(MemNode::Address); 2268 const Type *t2 = phase->type( adr ); 2269 if( t2 == Type::TOP ) return Type::TOP; 2270 const TypePtr *tp = t2->is_ptr(); 2271 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; 2272 const TypeAryPtr *tap = tp->isa_aryptr(); 2273 if( !tap ) return _type; 2274 return tap->size(); 2275 } 2276 2277 //-------------------------------Ideal--------------------------------------- 2278 // Feed through the length in AllocateArray(...length...)._length. 2279 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2280 Node* p = MemNode::Ideal_common(phase, can_reshape); 2281 if (p) return (p == NodeSentinel) ? NULL : p; 2282 2283 // Take apart the address into an oop and and offset. 2284 // Return 'this' if we cannot. 2285 Node* adr = in(MemNode::Address); 2286 intptr_t offset = 0; 2287 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2288 if (base == NULL) return NULL; 2289 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2290 if (tary == NULL) return NULL; 2291 2292 // We can fetch the length directly through an AllocateArrayNode. 2293 // This works even if the length is not constant (clone or newArray). 2294 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2295 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2296 if (alloc != NULL) { 2297 Node* allocated_length = alloc->Ideal_length(); 2298 Node* len = alloc->make_ideal_length(tary, phase); 2299 if (allocated_length != len) { 2300 // New CastII improves on this. 2301 return len; 2302 } 2303 } 2304 } 2305 2306 return NULL; 2307 } 2308 2309 //------------------------------Identity--------------------------------------- 2310 // Feed through the length in AllocateArray(...length...)._length. 2311 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { 2312 Node* x = LoadINode::Identity(phase); 2313 if (x != this) return x; 2314 2315 // Take apart the address into an oop and and offset. 2316 // Return 'this' if we cannot. 2317 Node* adr = in(MemNode::Address); 2318 intptr_t offset = 0; 2319 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 2320 if (base == NULL) return this; 2321 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); 2322 if (tary == NULL) return this; 2323 2324 // We can fetch the length directly through an AllocateArrayNode. 2325 // This works even if the length is not constant (clone or newArray). 2326 if (offset == arrayOopDesc::length_offset_in_bytes()) { 2327 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); 2328 if (alloc != NULL) { 2329 Node* allocated_length = alloc->Ideal_length(); 2330 // Do not allow make_ideal_length to allocate a CastII node. 2331 Node* len = alloc->make_ideal_length(tary, phase, false); 2332 if (allocated_length == len) { 2333 // Return allocated_length only if it would not be improved by a CastII. 2334 return allocated_length; 2335 } 2336 } 2337 } 2338 2339 return this; 2340 2341 } 2342 2343 //============================================================================= 2344 //---------------------------StoreNode::make----------------------------------- 2345 // Polymorphic factory method: 2346 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { 2347 assert((mo == unordered || mo == release), "unexpected"); 2348 Compile* C = gvn.C; 2349 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || 2350 ctl != NULL, "raw memory operations should have control edge"); 2351 2352 switch (bt) { 2353 case T_BOOLEAN: 2354 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo); 2355 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo); 2356 case T_CHAR: 2357 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo); 2358 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo); 2359 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo); 2360 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo); 2361 case T_METADATA: 2362 case T_ADDRESS: 2363 case T_OBJECT: 2364 #ifdef _LP64 2365 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2366 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop())); 2367 return new StoreNNode(ctl, mem, adr, adr_type, val, mo); 2368 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || 2369 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && 2370 adr->bottom_type()->isa_rawptr())) { 2371 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); 2372 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); 2373 } 2374 #endif 2375 { 2376 return new StorePNode(ctl, mem, adr, adr_type, val, mo); 2377 } 2378 } 2379 ShouldNotReachHere(); 2380 return (StoreNode*)NULL; 2381 } 2382 2383 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2384 bool require_atomic = true; 2385 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2386 } 2387 2388 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { 2389 bool require_atomic = true; 2390 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic); 2391 } 2392 2393 2394 //--------------------------bottom_type---------------------------------------- 2395 const Type *StoreNode::bottom_type() const { 2396 return Type::MEMORY; 2397 } 2398 2399 //------------------------------hash------------------------------------------- 2400 uint StoreNode::hash() const { 2401 // unroll addition of interesting fields 2402 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); 2403 2404 // Since they are not commoned, do not hash them: 2405 return NO_HASH; 2406 } 2407 2408 //------------------------------Ideal------------------------------------------ 2409 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). 2410 // When a store immediately follows a relevant allocation/initialization, 2411 // try to capture it into the initialization, or hoist it above. 2412 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2413 Node* p = MemNode::Ideal_common(phase, can_reshape); 2414 if (p) return (p == NodeSentinel) ? NULL : p; 2415 2416 Node* mem = in(MemNode::Memory); 2417 Node* address = in(MemNode::Address); 2418 2419 // Back-to-back stores to same address? Fold em up. Generally 2420 // unsafe if I have intervening uses... Also disallowed for StoreCM 2421 // since they must follow each StoreP operation. Redundant StoreCMs 2422 // are eliminated just before matching in final_graph_reshape. 2423 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && 2424 mem->Opcode() != Op_StoreCM) { 2425 // Looking at a dead closed cycle of memory? 2426 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); 2427 2428 assert(Opcode() == mem->Opcode() || 2429 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, 2430 "no mismatched stores, except on raw memory"); 2431 2432 if (mem->outcnt() == 1 && // check for intervening uses 2433 mem->as_Store()->memory_size() <= this->memory_size()) { 2434 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. 2435 // For example, 'mem' might be the final state at a conditional return. 2436 // Or, 'mem' might be used by some node which is live at the same time 2437 // 'this' is live, which might be unschedulable. So, require exactly 2438 // ONE user, the 'this' store, until such time as we clone 'mem' for 2439 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). 2440 if (can_reshape) { // (%%% is this an anachronism?) 2441 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), 2442 phase->is_IterGVN()); 2443 } else { 2444 // It's OK to do this in the parser, since DU info is always accurate, 2445 // and the parser always refers to nodes via SafePointNode maps. 2446 set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2447 } 2448 return this; 2449 } 2450 } 2451 2452 // Capture an unaliased, unconditional, simple store into an initializer. 2453 // Or, if it is independent of the allocation, hoist it above the allocation. 2454 if (ReduceFieldZeroing && /*can_reshape &&*/ 2455 mem->is_Proj() && mem->in(0)->is_Initialize()) { 2456 InitializeNode* init = mem->in(0)->as_Initialize(); 2457 intptr_t offset = init->can_capture_store(this, phase, can_reshape); 2458 if (offset > 0) { 2459 Node* moved = init->capture_store(this, offset, phase, can_reshape); 2460 // If the InitializeNode captured me, it made a raw copy of me, 2461 // and I need to disappear. 2462 if (moved != NULL) { 2463 // %%% hack to ensure that Ideal returns a new node: 2464 mem = MergeMemNode::make(phase->C, mem); 2465 return mem; // fold me away 2466 } 2467 } 2468 } 2469 2470 return NULL; // No further progress 2471 } 2472 2473 //------------------------------Value----------------------------------------- 2474 const Type *StoreNode::Value( PhaseTransform *phase ) const { 2475 // Either input is TOP ==> the result is TOP 2476 const Type *t1 = phase->type( in(MemNode::Memory) ); 2477 if( t1 == Type::TOP ) return Type::TOP; 2478 const Type *t2 = phase->type( in(MemNode::Address) ); 2479 if( t2 == Type::TOP ) return Type::TOP; 2480 const Type *t3 = phase->type( in(MemNode::ValueIn) ); 2481 if( t3 == Type::TOP ) return Type::TOP; 2482 return Type::MEMORY; 2483 } 2484 2485 //------------------------------Identity--------------------------------------- 2486 // Remove redundant stores: 2487 // Store(m, p, Load(m, p)) changes to m. 2488 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). 2489 Node *StoreNode::Identity( PhaseTransform *phase ) { 2490 Node* mem = in(MemNode::Memory); 2491 Node* adr = in(MemNode::Address); 2492 Node* val = in(MemNode::ValueIn); 2493 2494 // Load then Store? Then the Store is useless 2495 if (val->is_Load() && 2496 val->in(MemNode::Address)->eqv_uncast(adr) && 2497 val->in(MemNode::Memory )->eqv_uncast(mem) && 2498 val->as_Load()->store_Opcode() == Opcode()) { 2499 return mem; 2500 } 2501 2502 // Two stores in a row of the same value? 2503 if (mem->is_Store() && 2504 mem->in(MemNode::Address)->eqv_uncast(adr) && 2505 mem->in(MemNode::ValueIn)->eqv_uncast(val) && 2506 mem->Opcode() == Opcode()) { 2507 return mem; 2508 } 2509 2510 // Store of zero anywhere into a freshly-allocated object? 2511 // Then the store is useless. 2512 // (It must already have been captured by the InitializeNode.) 2513 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { 2514 // a newly allocated object is already all-zeroes everywhere 2515 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { 2516 return mem; 2517 } 2518 2519 // the store may also apply to zero-bits in an earlier object 2520 Node* prev_mem = find_previous_store(phase); 2521 // Steps (a), (b): Walk past independent stores to find an exact match. 2522 if (prev_mem != NULL) { 2523 Node* prev_val = can_see_stored_value(prev_mem, phase); 2524 if (prev_val != NULL && phase->eqv(prev_val, val)) { 2525 // prev_val and val might differ by a cast; it would be good 2526 // to keep the more informative of the two. 2527 return mem; 2528 } 2529 } 2530 } 2531 2532 return this; 2533 } 2534 2535 //------------------------------match_edge------------------------------------- 2536 // Do we Match on this edge index or not? Match only memory & value 2537 uint StoreNode::match_edge(uint idx) const { 2538 return idx == MemNode::Address || idx == MemNode::ValueIn; 2539 } 2540 2541 //------------------------------cmp-------------------------------------------- 2542 // Do not common stores up together. They generally have to be split 2543 // back up anyways, so do not bother. 2544 uint StoreNode::cmp( const Node &n ) const { 2545 return (&n == this); // Always fail except on self 2546 } 2547 2548 //------------------------------Ideal_masked_input----------------------------- 2549 // Check for a useless mask before a partial-word store 2550 // (StoreB ... (AndI valIn conIa) ) 2551 // If (conIa & mask == mask) this simplifies to 2552 // (StoreB ... (valIn) ) 2553 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { 2554 Node *val = in(MemNode::ValueIn); 2555 if( val->Opcode() == Op_AndI ) { 2556 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2557 if( t && t->is_con() && (t->get_con() & mask) == mask ) { 2558 set_req(MemNode::ValueIn, val->in(1)); 2559 return this; 2560 } 2561 } 2562 return NULL; 2563 } 2564 2565 2566 //------------------------------Ideal_sign_extended_input---------------------- 2567 // Check for useless sign-extension before a partial-word store 2568 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) 2569 // If (conIL == conIR && conIR <= num_bits) this simplifies to 2570 // (StoreB ... (valIn) ) 2571 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { 2572 Node *val = in(MemNode::ValueIn); 2573 if( val->Opcode() == Op_RShiftI ) { 2574 const TypeInt *t = phase->type( val->in(2) )->isa_int(); 2575 if( t && t->is_con() && (t->get_con() <= num_bits) ) { 2576 Node *shl = val->in(1); 2577 if( shl->Opcode() == Op_LShiftI ) { 2578 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); 2579 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { 2580 set_req(MemNode::ValueIn, shl->in(1)); 2581 return this; 2582 } 2583 } 2584 } 2585 } 2586 return NULL; 2587 } 2588 2589 //------------------------------value_never_loaded----------------------------------- 2590 // Determine whether there are any possible loads of the value stored. 2591 // For simplicity, we actually check if there are any loads from the 2592 // address stored to, not just for loads of the value stored by this node. 2593 // 2594 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { 2595 Node *adr = in(Address); 2596 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); 2597 if (adr_oop == NULL) 2598 return false; 2599 if (!adr_oop->is_known_instance_field()) 2600 return false; // if not a distinct instance, there may be aliases of the address 2601 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { 2602 Node *use = adr->fast_out(i); 2603 int opc = use->Opcode(); 2604 if (use->is_Load() || use->is_LoadStore()) { 2605 return false; 2606 } 2607 } 2608 return true; 2609 } 2610 2611 //============================================================================= 2612 //------------------------------Ideal------------------------------------------ 2613 // If the store is from an AND mask that leaves the low bits untouched, then 2614 // we can skip the AND operation. If the store is from a sign-extension 2615 // (a left shift, then right shift) we can skip both. 2616 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2617 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); 2618 if( progress != NULL ) return progress; 2619 2620 progress = StoreNode::Ideal_sign_extended_input(phase, 24); 2621 if( progress != NULL ) return progress; 2622 2623 // Finally check the default case 2624 return StoreNode::Ideal(phase, can_reshape); 2625 } 2626 2627 //============================================================================= 2628 //------------------------------Ideal------------------------------------------ 2629 // If the store is from an AND mask that leaves the low bits untouched, then 2630 // we can skip the AND operation 2631 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2632 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); 2633 if( progress != NULL ) return progress; 2634 2635 progress = StoreNode::Ideal_sign_extended_input(phase, 16); 2636 if( progress != NULL ) return progress; 2637 2638 // Finally check the default case 2639 return StoreNode::Ideal(phase, can_reshape); 2640 } 2641 2642 //============================================================================= 2643 //------------------------------Identity--------------------------------------- 2644 Node *StoreCMNode::Identity( PhaseTransform *phase ) { 2645 // No need to card mark when storing a null ptr 2646 Node* my_store = in(MemNode::OopStore); 2647 if (my_store->is_Store()) { 2648 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); 2649 if( t1 == TypePtr::NULL_PTR ) { 2650 return in(MemNode::Memory); 2651 } 2652 } 2653 return this; 2654 } 2655 2656 //============================================================================= 2657 //------------------------------Ideal--------------------------------------- 2658 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2659 Node* progress = StoreNode::Ideal(phase, can_reshape); 2660 if (progress != NULL) return progress; 2661 2662 Node* my_store = in(MemNode::OopStore); 2663 if (my_store->is_MergeMem()) { 2664 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); 2665 set_req(MemNode::OopStore, mem); 2666 return this; 2667 } 2668 2669 return NULL; 2670 } 2671 2672 //------------------------------Value----------------------------------------- 2673 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { 2674 // Either input is TOP ==> the result is TOP 2675 const Type *t = phase->type( in(MemNode::Memory) ); 2676 if( t == Type::TOP ) return Type::TOP; 2677 t = phase->type( in(MemNode::Address) ); 2678 if( t == Type::TOP ) return Type::TOP; 2679 t = phase->type( in(MemNode::ValueIn) ); 2680 if( t == Type::TOP ) return Type::TOP; 2681 // If extra input is TOP ==> the result is TOP 2682 t = phase->type( in(MemNode::OopStore) ); 2683 if( t == Type::TOP ) return Type::TOP; 2684 2685 return StoreNode::Value( phase ); 2686 } 2687 2688 2689 //============================================================================= 2690 //----------------------------------SCMemProjNode------------------------------ 2691 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const 2692 { 2693 return bottom_type(); 2694 } 2695 2696 //============================================================================= 2697 //----------------------------------LoadStoreNode------------------------------ 2698 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) 2699 : Node(required), 2700 _type(rt), 2701 _adr_type(at) 2702 { 2703 init_req(MemNode::Control, c ); 2704 init_req(MemNode::Memory , mem); 2705 init_req(MemNode::Address, adr); 2706 init_req(MemNode::ValueIn, val); 2707 init_class_id(Class_LoadStore); 2708 } 2709 2710 uint LoadStoreNode::ideal_reg() const { 2711 return _type->ideal_reg(); 2712 } 2713 2714 bool LoadStoreNode::result_not_used() const { 2715 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { 2716 Node *x = fast_out(i); 2717 if (x->Opcode() == Op_SCMemProj) continue; 2718 return false; 2719 } 2720 return true; 2721 } 2722 2723 uint LoadStoreNode::size_of() const { return sizeof(*this); } 2724 2725 //============================================================================= 2726 //----------------------------------LoadStoreConditionalNode-------------------- 2727 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { 2728 init_req(ExpectedIn, ex ); 2729 } 2730 2731 //============================================================================= 2732 //-------------------------------adr_type-------------------------------------- 2733 // Do we Match on this edge index or not? Do not match memory 2734 const TypePtr* ClearArrayNode::adr_type() const { 2735 Node *adr = in(3); 2736 return MemNode::calculate_adr_type(adr->bottom_type()); 2737 } 2738 2739 //------------------------------match_edge------------------------------------- 2740 // Do we Match on this edge index or not? Do not match memory 2741 uint ClearArrayNode::match_edge(uint idx) const { 2742 return idx > 1; 2743 } 2744 2745 //------------------------------Identity--------------------------------------- 2746 // Clearing a zero length array does nothing 2747 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { 2748 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; 2749 } 2750 2751 //------------------------------Idealize--------------------------------------- 2752 // Clearing a short array is faster with stores 2753 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ 2754 const int unit = BytesPerLong; 2755 const TypeX* t = phase->type(in(2))->isa_intptr_t(); 2756 if (!t) return NULL; 2757 if (!t->is_con()) return NULL; 2758 intptr_t raw_count = t->get_con(); 2759 intptr_t size = raw_count; 2760 if (!Matcher::init_array_count_is_in_bytes) size *= unit; 2761 // Clearing nothing uses the Identity call. 2762 // Negative clears are possible on dead ClearArrays 2763 // (see jck test stmt114.stmt11402.val). 2764 if (size <= 0 || size % unit != 0) return NULL; 2765 intptr_t count = size / unit; 2766 // Length too long; use fast hardware clear 2767 if (size > Matcher::init_array_short_size) return NULL; 2768 Node *mem = in(1); 2769 if( phase->type(mem)==Type::TOP ) return NULL; 2770 Node *adr = in(3); 2771 const Type* at = phase->type(adr); 2772 if( at==Type::TOP ) return NULL; 2773 const TypePtr* atp = at->isa_ptr(); 2774 // adjust atp to be the correct array element address type 2775 if (atp == NULL) atp = TypePtr::BOTTOM; 2776 else atp = atp->add_offset(Type::OffsetBot); 2777 // Get base for derived pointer purposes 2778 if( adr->Opcode() != Op_AddP ) Unimplemented(); 2779 Node *base = adr->in(1); 2780 2781 Node *zero = phase->makecon(TypeLong::ZERO); 2782 Node *off = phase->MakeConX(BytesPerLong); 2783 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2784 count--; 2785 while( count-- ) { 2786 mem = phase->transform(mem); 2787 adr = phase->transform(new AddPNode(base,adr,off)); 2788 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); 2789 } 2790 return mem; 2791 } 2792 2793 //----------------------------step_through---------------------------------- 2794 // Return allocation input memory edge if it is different instance 2795 // or itself if it is the one we are looking for. 2796 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { 2797 Node* n = *np; 2798 assert(n->is_ClearArray(), "sanity"); 2799 intptr_t offset; 2800 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); 2801 // This method is called only before Allocate nodes are expanded during 2802 // macro nodes expansion. Before that ClearArray nodes are only generated 2803 // in LibraryCallKit::generate_arraycopy() which follows allocations. 2804 assert(alloc != NULL, "should have allocation"); 2805 if (alloc->_idx == instance_id) { 2806 // Can not bypass initialization of the instance we are looking for. 2807 return false; 2808 } 2809 // Otherwise skip it. 2810 InitializeNode* init = alloc->initialization(); 2811 if (init != NULL) 2812 *np = init->in(TypeFunc::Memory); 2813 else 2814 *np = alloc->in(TypeFunc::Memory); 2815 return true; 2816 } 2817 2818 //----------------------------clear_memory------------------------------------- 2819 // Generate code to initialize object storage to zero. 2820 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2821 intptr_t start_offset, 2822 Node* end_offset, 2823 PhaseGVN* phase) { 2824 Compile* C = phase->C; 2825 intptr_t offset = start_offset; 2826 2827 int unit = BytesPerLong; 2828 if ((offset % unit) != 0) { 2829 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset)); 2830 adr = phase->transform(adr); 2831 const TypePtr* atp = TypeRawPtr::BOTTOM; 2832 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2833 mem = phase->transform(mem); 2834 offset += BytesPerInt; 2835 } 2836 assert((offset % unit) == 0, ""); 2837 2838 // Initialize the remaining stuff, if any, with a ClearArray. 2839 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); 2840 } 2841 2842 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2843 Node* start_offset, 2844 Node* end_offset, 2845 PhaseGVN* phase) { 2846 if (start_offset == end_offset) { 2847 // nothing to do 2848 return mem; 2849 } 2850 2851 Compile* C = phase->C; 2852 int unit = BytesPerLong; 2853 Node* zbase = start_offset; 2854 Node* zend = end_offset; 2855 2856 // Scale to the unit required by the CPU: 2857 if (!Matcher::init_array_count_is_in_bytes) { 2858 Node* shift = phase->intcon(exact_log2(unit)); 2859 zbase = phase->transform(new URShiftXNode(zbase, shift) ); 2860 zend = phase->transform(new URShiftXNode(zend, shift) ); 2861 } 2862 2863 // Bulk clear double-words 2864 Node* zsize = phase->transform(new SubXNode(zend, zbase) ); 2865 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) ); 2866 mem = new ClearArrayNode(ctl, mem, zsize, adr); 2867 return phase->transform(mem); 2868 } 2869 2870 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, 2871 intptr_t start_offset, 2872 intptr_t end_offset, 2873 PhaseGVN* phase) { 2874 if (start_offset == end_offset) { 2875 // nothing to do 2876 return mem; 2877 } 2878 2879 Compile* C = phase->C; 2880 assert((end_offset % BytesPerInt) == 0, "odd end offset"); 2881 intptr_t done_offset = end_offset; 2882 if ((done_offset % BytesPerLong) != 0) { 2883 done_offset -= BytesPerInt; 2884 } 2885 if (done_offset > start_offset) { 2886 mem = clear_memory(ctl, mem, dest, 2887 start_offset, phase->MakeConX(done_offset), phase); 2888 } 2889 if (done_offset < end_offset) { // emit the final 32-bit store 2890 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset)); 2891 adr = phase->transform(adr); 2892 const TypePtr* atp = TypeRawPtr::BOTTOM; 2893 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); 2894 mem = phase->transform(mem); 2895 done_offset += BytesPerInt; 2896 } 2897 assert(done_offset == end_offset, ""); 2898 return mem; 2899 } 2900 2901 //============================================================================= 2902 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) 2903 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), 2904 _adr_type(C->get_adr_type(alias_idx)) 2905 { 2906 init_class_id(Class_MemBar); 2907 Node* top = C->top(); 2908 init_req(TypeFunc::I_O,top); 2909 init_req(TypeFunc::FramePtr,top); 2910 init_req(TypeFunc::ReturnAdr,top); 2911 if (precedent != NULL) 2912 init_req(TypeFunc::Parms, precedent); 2913 } 2914 2915 //------------------------------cmp-------------------------------------------- 2916 uint MemBarNode::hash() const { return NO_HASH; } 2917 uint MemBarNode::cmp( const Node &n ) const { 2918 return (&n == this); // Always fail except on self 2919 } 2920 2921 //------------------------------make------------------------------------------- 2922 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { 2923 switch (opcode) { 2924 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn); 2925 case Op_LoadFence: return new LoadFenceNode(C, atp, pn); 2926 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn); 2927 case Op_StoreFence: return new StoreFenceNode(C, atp, pn); 2928 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn); 2929 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn); 2930 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn); 2931 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn); 2932 case Op_Initialize: return new InitializeNode(C, atp, pn); 2933 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn); 2934 default: ShouldNotReachHere(); return NULL; 2935 } 2936 } 2937 2938 //------------------------------Ideal------------------------------------------ 2939 // Return a node which is more "ideal" than the current node. Strip out 2940 // control copies 2941 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2942 if (remove_dead_region(phase, can_reshape)) return this; 2943 // Don't bother trying to transform a dead node 2944 if (in(0) && in(0)->is_top()) { 2945 return NULL; 2946 } 2947 2948 // Eliminate volatile MemBars for scalar replaced objects. 2949 if (can_reshape && req() == (Precedent+1)) { 2950 bool eliminate = false; 2951 int opc = Opcode(); 2952 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { 2953 // Volatile field loads and stores. 2954 Node* my_mem = in(MemBarNode::Precedent); 2955 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge 2956 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { 2957 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, 2958 // replace this Precedent (decodeN) with the Load instead. 2959 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { 2960 Node* load_node = my_mem->in(1); 2961 set_req(MemBarNode::Precedent, load_node); 2962 phase->is_IterGVN()->_worklist.push(my_mem); 2963 my_mem = load_node; 2964 } else { 2965 assert(my_mem->unique_out() == this, "sanity"); 2966 del_req(Precedent); 2967 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later 2968 my_mem = NULL; 2969 } 2970 } 2971 if (my_mem != NULL && my_mem->is_Mem()) { 2972 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); 2973 // Check for scalar replaced object reference. 2974 if( t_oop != NULL && t_oop->is_known_instance_field() && 2975 t_oop->offset() != Type::OffsetBot && 2976 t_oop->offset() != Type::OffsetTop) { 2977 eliminate = true; 2978 } 2979 } 2980 } else if (opc == Op_MemBarRelease) { 2981 // Final field stores. 2982 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); 2983 if ((alloc != NULL) && alloc->is_Allocate() && 2984 alloc->as_Allocate()->_is_non_escaping) { 2985 // The allocated object does not escape. 2986 eliminate = true; 2987 } 2988 } 2989 if (eliminate) { 2990 // Replace MemBar projections by its inputs. 2991 PhaseIterGVN* igvn = phase->is_IterGVN(); 2992 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); 2993 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); 2994 // Must return either the original node (now dead) or a new node 2995 // (Do not return a top here, since that would break the uniqueness of top.) 2996 return new ConINode(TypeInt::ZERO); 2997 } 2998 } 2999 return NULL; 3000 } 3001 3002 //------------------------------Value------------------------------------------ 3003 const Type *MemBarNode::Value( PhaseTransform *phase ) const { 3004 if( !in(0) ) return Type::TOP; 3005 if( phase->type(in(0)) == Type::TOP ) 3006 return Type::TOP; 3007 return TypeTuple::MEMBAR; 3008 } 3009 3010 //------------------------------match------------------------------------------ 3011 // Construct projections for memory. 3012 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { 3013 switch (proj->_con) { 3014 case TypeFunc::Control: 3015 case TypeFunc::Memory: 3016 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 3017 } 3018 ShouldNotReachHere(); 3019 return NULL; 3020 } 3021 3022 //===========================InitializeNode==================================== 3023 // SUMMARY: 3024 // This node acts as a memory barrier on raw memory, after some raw stores. 3025 // The 'cooked' oop value feeds from the Initialize, not the Allocation. 3026 // The Initialize can 'capture' suitably constrained stores as raw inits. 3027 // It can coalesce related raw stores into larger units (called 'tiles'). 3028 // It can avoid zeroing new storage for memory units which have raw inits. 3029 // At macro-expansion, it is marked 'complete', and does not optimize further. 3030 // 3031 // EXAMPLE: 3032 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. 3033 // ctl = incoming control; mem* = incoming memory 3034 // (Note: A star * on a memory edge denotes I/O and other standard edges.) 3035 // First allocate uninitialized memory and fill in the header: 3036 // alloc = (Allocate ctl mem* 16 #short[].klass ...) 3037 // ctl := alloc.Control; mem* := alloc.Memory* 3038 // rawmem = alloc.Memory; rawoop = alloc.RawAddress 3039 // Then initialize to zero the non-header parts of the raw memory block: 3040 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) 3041 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory 3042 // After the initialize node executes, the object is ready for service: 3043 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) 3044 // Suppose its body is immediately initialized as {1,2}: 3045 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3046 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3047 // mem.SLICE(#short[*]) := store2 3048 // 3049 // DETAILS: 3050 // An InitializeNode collects and isolates object initialization after 3051 // an AllocateNode and before the next possible safepoint. As a 3052 // memory barrier (MemBarNode), it keeps critical stores from drifting 3053 // down past any safepoint or any publication of the allocation. 3054 // Before this barrier, a newly-allocated object may have uninitialized bits. 3055 // After this barrier, it may be treated as a real oop, and GC is allowed. 3056 // 3057 // The semantics of the InitializeNode include an implicit zeroing of 3058 // the new object from object header to the end of the object. 3059 // (The object header and end are determined by the AllocateNode.) 3060 // 3061 // Certain stores may be added as direct inputs to the InitializeNode. 3062 // These stores must update raw memory, and they must be to addresses 3063 // derived from the raw address produced by AllocateNode, and with 3064 // a constant offset. They must be ordered by increasing offset. 3065 // The first one is at in(RawStores), the last at in(req()-1). 3066 // Unlike most memory operations, they are not linked in a chain, 3067 // but are displayed in parallel as users of the rawmem output of 3068 // the allocation. 3069 // 3070 // (See comments in InitializeNode::capture_store, which continue 3071 // the example given above.) 3072 // 3073 // When the associated Allocate is macro-expanded, the InitializeNode 3074 // may be rewritten to optimize collected stores. A ClearArrayNode 3075 // may also be created at that point to represent any required zeroing. 3076 // The InitializeNode is then marked 'complete', prohibiting further 3077 // capturing of nearby memory operations. 3078 // 3079 // During macro-expansion, all captured initializations which store 3080 // constant values of 32 bits or smaller are coalesced (if advantageous) 3081 // into larger 'tiles' 32 or 64 bits. This allows an object to be 3082 // initialized in fewer memory operations. Memory words which are 3083 // covered by neither tiles nor non-constant stores are pre-zeroed 3084 // by explicit stores of zero. (The code shape happens to do all 3085 // zeroing first, then all other stores, with both sequences occurring 3086 // in order of ascending offsets.) 3087 // 3088 // Alternatively, code may be inserted between an AllocateNode and its 3089 // InitializeNode, to perform arbitrary initialization of the new object. 3090 // E.g., the object copying intrinsics insert complex data transfers here. 3091 // The initialization must then be marked as 'complete' disable the 3092 // built-in zeroing semantics and the collection of initializing stores. 3093 // 3094 // While an InitializeNode is incomplete, reads from the memory state 3095 // produced by it are optimizable if they match the control edge and 3096 // new oop address associated with the allocation/initialization. 3097 // They return a stored value (if the offset matches) or else zero. 3098 // A write to the memory state, if it matches control and address, 3099 // and if it is to a constant offset, may be 'captured' by the 3100 // InitializeNode. It is cloned as a raw memory operation and rewired 3101 // inside the initialization, to the raw oop produced by the allocation. 3102 // Operations on addresses which are provably distinct (e.g., to 3103 // other AllocateNodes) are allowed to bypass the initialization. 3104 // 3105 // The effect of all this is to consolidate object initialization 3106 // (both arrays and non-arrays, both piecewise and bulk) into a 3107 // single location, where it can be optimized as a unit. 3108 // 3109 // Only stores with an offset less than TrackedInitializationLimit words 3110 // will be considered for capture by an InitializeNode. This puts a 3111 // reasonable limit on the complexity of optimized initializations. 3112 3113 //---------------------------InitializeNode------------------------------------ 3114 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) 3115 : _is_complete(Incomplete), _does_not_escape(false), 3116 MemBarNode(C, adr_type, rawoop) 3117 { 3118 init_class_id(Class_Initialize); 3119 3120 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); 3121 assert(in(RawAddress) == rawoop, "proper init"); 3122 // Note: allocation() can be NULL, for secondary initialization barriers 3123 } 3124 3125 // Since this node is not matched, it will be processed by the 3126 // register allocator. Declare that there are no constraints 3127 // on the allocation of the RawAddress edge. 3128 const RegMask &InitializeNode::in_RegMask(uint idx) const { 3129 // This edge should be set to top, by the set_complete. But be conservative. 3130 if (idx == InitializeNode::RawAddress) 3131 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); 3132 return RegMask::Empty; 3133 } 3134 3135 Node* InitializeNode::memory(uint alias_idx) { 3136 Node* mem = in(Memory); 3137 if (mem->is_MergeMem()) { 3138 return mem->as_MergeMem()->memory_at(alias_idx); 3139 } else { 3140 // incoming raw memory is not split 3141 return mem; 3142 } 3143 } 3144 3145 bool InitializeNode::is_non_zero() { 3146 if (is_complete()) return false; 3147 remove_extra_zeroes(); 3148 return (req() > RawStores); 3149 } 3150 3151 void InitializeNode::set_complete(PhaseGVN* phase) { 3152 assert(!is_complete(), "caller responsibility"); 3153 _is_complete = Complete; 3154 3155 // After this node is complete, it contains a bunch of 3156 // raw-memory initializations. There is no need for 3157 // it to have anything to do with non-raw memory effects. 3158 // Therefore, tell all non-raw users to re-optimize themselves, 3159 // after skipping the memory effects of this initialization. 3160 PhaseIterGVN* igvn = phase->is_IterGVN(); 3161 if (igvn) igvn->add_users_to_worklist(this); 3162 } 3163 3164 // convenience function 3165 // return false if the init contains any stores already 3166 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { 3167 InitializeNode* init = initialization(); 3168 if (init == NULL || init->is_complete()) return false; 3169 init->remove_extra_zeroes(); 3170 // for now, if this allocation has already collected any inits, bail: 3171 if (init->is_non_zero()) return false; 3172 init->set_complete(phase); 3173 return true; 3174 } 3175 3176 void InitializeNode::remove_extra_zeroes() { 3177 if (req() == RawStores) return; 3178 Node* zmem = zero_memory(); 3179 uint fill = RawStores; 3180 for (uint i = fill; i < req(); i++) { 3181 Node* n = in(i); 3182 if (n->is_top() || n == zmem) continue; // skip 3183 if (fill < i) set_req(fill, n); // compact 3184 ++fill; 3185 } 3186 // delete any empty spaces created: 3187 while (fill < req()) { 3188 del_req(fill); 3189 } 3190 } 3191 3192 // Helper for remembering which stores go with which offsets. 3193 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { 3194 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node 3195 intptr_t offset = -1; 3196 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), 3197 phase, offset); 3198 if (base == NULL) return -1; // something is dead, 3199 if (offset < 0) return -1; // dead, dead 3200 return offset; 3201 } 3202 3203 // Helper for proving that an initialization expression is 3204 // "simple enough" to be folded into an object initialization. 3205 // Attempts to prove that a store's initial value 'n' can be captured 3206 // within the initialization without creating a vicious cycle, such as: 3207 // { Foo p = new Foo(); p.next = p; } 3208 // True for constants and parameters and small combinations thereof. 3209 bool InitializeNode::detect_init_independence(Node* n, int& count) { 3210 if (n == NULL) return true; // (can this really happen?) 3211 if (n->is_Proj()) n = n->in(0); 3212 if (n == this) return false; // found a cycle 3213 if (n->is_Con()) return true; 3214 if (n->is_Start()) return true; // params, etc., are OK 3215 if (n->is_Root()) return true; // even better 3216 3217 Node* ctl = n->in(0); 3218 if (ctl != NULL && !ctl->is_top()) { 3219 if (ctl->is_Proj()) ctl = ctl->in(0); 3220 if (ctl == this) return false; 3221 3222 // If we already know that the enclosing memory op is pinned right after 3223 // the init, then any control flow that the store has picked up 3224 // must have preceded the init, or else be equal to the init. 3225 // Even after loop optimizations (which might change control edges) 3226 // a store is never pinned *before* the availability of its inputs. 3227 if (!MemNode::all_controls_dominate(n, this)) 3228 return false; // failed to prove a good control 3229 } 3230 3231 // Check data edges for possible dependencies on 'this'. 3232 if ((count += 1) > 20) return false; // complexity limit 3233 for (uint i = 1; i < n->req(); i++) { 3234 Node* m = n->in(i); 3235 if (m == NULL || m == n || m->is_top()) continue; 3236 uint first_i = n->find_edge(m); 3237 if (i != first_i) continue; // process duplicate edge just once 3238 if (!detect_init_independence(m, count)) { 3239 return false; 3240 } 3241 } 3242 3243 return true; 3244 } 3245 3246 // Here are all the checks a Store must pass before it can be moved into 3247 // an initialization. Returns zero if a check fails. 3248 // On success, returns the (constant) offset to which the store applies, 3249 // within the initialized memory. 3250 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { 3251 const int FAIL = 0; 3252 if (st->req() != MemNode::ValueIn + 1) 3253 return FAIL; // an inscrutable StoreNode (card mark?) 3254 Node* ctl = st->in(MemNode::Control); 3255 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) 3256 return FAIL; // must be unconditional after the initialization 3257 Node* mem = st->in(MemNode::Memory); 3258 if (!(mem->is_Proj() && mem->in(0) == this)) 3259 return FAIL; // must not be preceded by other stores 3260 Node* adr = st->in(MemNode::Address); 3261 intptr_t offset; 3262 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); 3263 if (alloc == NULL) 3264 return FAIL; // inscrutable address 3265 if (alloc != allocation()) 3266 return FAIL; // wrong allocation! (store needs to float up) 3267 Node* val = st->in(MemNode::ValueIn); 3268 int complexity_count = 0; 3269 if (!detect_init_independence(val, complexity_count)) 3270 return FAIL; // stored value must be 'simple enough' 3271 3272 // The Store can be captured only if nothing after the allocation 3273 // and before the Store is using the memory location that the store 3274 // overwrites. 3275 bool failed = false; 3276 // If is_complete_with_arraycopy() is true the shape of the graph is 3277 // well defined and is safe so no need for extra checks. 3278 if (!is_complete_with_arraycopy()) { 3279 // We are going to look at each use of the memory state following 3280 // the allocation to make sure nothing reads the memory that the 3281 // Store writes. 3282 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); 3283 int alias_idx = phase->C->get_alias_index(t_adr); 3284 ResourceMark rm; 3285 Unique_Node_List mems; 3286 mems.push(mem); 3287 Node* unique_merge = NULL; 3288 for (uint next = 0; next < mems.size(); ++next) { 3289 Node *m = mems.at(next); 3290 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3291 Node *n = m->fast_out(j); 3292 if (n->outcnt() == 0) { 3293 continue; 3294 } 3295 if (n == st) { 3296 continue; 3297 } else if (n->in(0) != NULL && n->in(0) != ctl) { 3298 // If the control of this use is different from the control 3299 // of the Store which is right after the InitializeNode then 3300 // this node cannot be between the InitializeNode and the 3301 // Store. 3302 continue; 3303 } else if (n->is_MergeMem()) { 3304 if (n->as_MergeMem()->memory_at(alias_idx) == m) { 3305 // We can hit a MergeMemNode (that will likely go away 3306 // later) that is a direct use of the memory state 3307 // following the InitializeNode on the same slice as the 3308 // store node that we'd like to capture. We need to check 3309 // the uses of the MergeMemNode. 3310 mems.push(n); 3311 } 3312 } else if (n->is_Mem()) { 3313 Node* other_adr = n->in(MemNode::Address); 3314 if (other_adr == adr) { 3315 failed = true; 3316 break; 3317 } else { 3318 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); 3319 if (other_t_adr != NULL) { 3320 int other_alias_idx = phase->C->get_alias_index(other_t_adr); 3321 if (other_alias_idx == alias_idx) { 3322 // A load from the same memory slice as the store right 3323 // after the InitializeNode. We check the control of the 3324 // object/array that is loaded from. If it's the same as 3325 // the store control then we cannot capture the store. 3326 assert(!n->is_Store(), "2 stores to same slice on same control?"); 3327 Node* base = other_adr; 3328 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); 3329 base = base->in(AddPNode::Base); 3330 if (base != NULL) { 3331 base = base->uncast(); 3332 if (base->is_Proj() && base->in(0) == alloc) { 3333 failed = true; 3334 break; 3335 } 3336 } 3337 } 3338 } 3339 } 3340 } else { 3341 failed = true; 3342 break; 3343 } 3344 } 3345 } 3346 } 3347 if (failed) { 3348 if (!can_reshape) { 3349 // We decided we couldn't capture the store during parsing. We 3350 // should try again during the next IGVN once the graph is 3351 // cleaner. 3352 phase->C->record_for_igvn(st); 3353 } 3354 return FAIL; 3355 } 3356 3357 return offset; // success 3358 } 3359 3360 // Find the captured store in(i) which corresponds to the range 3361 // [start..start+size) in the initialized object. 3362 // If there is one, return its index i. If there isn't, return the 3363 // negative of the index where it should be inserted. 3364 // Return 0 if the queried range overlaps an initialization boundary 3365 // or if dead code is encountered. 3366 // If size_in_bytes is zero, do not bother with overlap checks. 3367 int InitializeNode::captured_store_insertion_point(intptr_t start, 3368 int size_in_bytes, 3369 PhaseTransform* phase) { 3370 const int FAIL = 0, MAX_STORE = BytesPerLong; 3371 3372 if (is_complete()) 3373 return FAIL; // arraycopy got here first; punt 3374 3375 assert(allocation() != NULL, "must be present"); 3376 3377 // no negatives, no header fields: 3378 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; 3379 3380 // after a certain size, we bail out on tracking all the stores: 3381 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3382 if (start >= ti_limit) return FAIL; 3383 3384 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { 3385 if (i >= limit) return -(int)i; // not found; here is where to put it 3386 3387 Node* st = in(i); 3388 intptr_t st_off = get_store_offset(st, phase); 3389 if (st_off < 0) { 3390 if (st != zero_memory()) { 3391 return FAIL; // bail out if there is dead garbage 3392 } 3393 } else if (st_off > start) { 3394 // ...we are done, since stores are ordered 3395 if (st_off < start + size_in_bytes) { 3396 return FAIL; // the next store overlaps 3397 } 3398 return -(int)i; // not found; here is where to put it 3399 } else if (st_off < start) { 3400 if (size_in_bytes != 0 && 3401 start < st_off + MAX_STORE && 3402 start < st_off + st->as_Store()->memory_size()) { 3403 return FAIL; // the previous store overlaps 3404 } 3405 } else { 3406 if (size_in_bytes != 0 && 3407 st->as_Store()->memory_size() != size_in_bytes) { 3408 return FAIL; // mismatched store size 3409 } 3410 return i; 3411 } 3412 3413 ++i; 3414 } 3415 } 3416 3417 // Look for a captured store which initializes at the offset 'start' 3418 // with the given size. If there is no such store, and no other 3419 // initialization interferes, then return zero_memory (the memory 3420 // projection of the AllocateNode). 3421 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, 3422 PhaseTransform* phase) { 3423 assert(stores_are_sane(phase), ""); 3424 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3425 if (i == 0) { 3426 return NULL; // something is dead 3427 } else if (i < 0) { 3428 return zero_memory(); // just primordial zero bits here 3429 } else { 3430 Node* st = in(i); // here is the store at this position 3431 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); 3432 return st; 3433 } 3434 } 3435 3436 // Create, as a raw pointer, an address within my new object at 'offset'. 3437 Node* InitializeNode::make_raw_address(intptr_t offset, 3438 PhaseTransform* phase) { 3439 Node* addr = in(RawAddress); 3440 if (offset != 0) { 3441 Compile* C = phase->C; 3442 addr = phase->transform( new AddPNode(C->top(), addr, 3443 phase->MakeConX(offset)) ); 3444 } 3445 return addr; 3446 } 3447 3448 // Clone the given store, converting it into a raw store 3449 // initializing a field or element of my new object. 3450 // Caller is responsible for retiring the original store, 3451 // with subsume_node or the like. 3452 // 3453 // From the example above InitializeNode::InitializeNode, 3454 // here are the old stores to be captured: 3455 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) 3456 // store2 = (StoreC init.Control store1 (+ oop 14) 2) 3457 // 3458 // Here is the changed code; note the extra edges on init: 3459 // alloc = (Allocate ...) 3460 // rawoop = alloc.RawAddress 3461 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) 3462 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) 3463 // init = (Initialize alloc.Control alloc.Memory rawoop 3464 // rawstore1 rawstore2) 3465 // 3466 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, 3467 PhaseTransform* phase, bool can_reshape) { 3468 assert(stores_are_sane(phase), ""); 3469 3470 if (start < 0) return NULL; 3471 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); 3472 3473 Compile* C = phase->C; 3474 int size_in_bytes = st->memory_size(); 3475 int i = captured_store_insertion_point(start, size_in_bytes, phase); 3476 if (i == 0) return NULL; // bail out 3477 Node* prev_mem = NULL; // raw memory for the captured store 3478 if (i > 0) { 3479 prev_mem = in(i); // there is a pre-existing store under this one 3480 set_req(i, C->top()); // temporarily disconnect it 3481 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. 3482 } else { 3483 i = -i; // no pre-existing store 3484 prev_mem = zero_memory(); // a slice of the newly allocated object 3485 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) 3486 set_req(--i, C->top()); // reuse this edge; it has been folded away 3487 else 3488 ins_req(i, C->top()); // build a new edge 3489 } 3490 Node* new_st = st->clone(); 3491 new_st->set_req(MemNode::Control, in(Control)); 3492 new_st->set_req(MemNode::Memory, prev_mem); 3493 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); 3494 new_st = phase->transform(new_st); 3495 3496 // At this point, new_st might have swallowed a pre-existing store 3497 // at the same offset, or perhaps new_st might have disappeared, 3498 // if it redundantly stored the same value (or zero to fresh memory). 3499 3500 // In any case, wire it in: 3501 set_req(i, new_st); 3502 3503 // The caller may now kill the old guy. 3504 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); 3505 assert(check_st == new_st || check_st == NULL, "must be findable"); 3506 assert(!is_complete(), ""); 3507 return new_st; 3508 } 3509 3510 static bool store_constant(jlong* tiles, int num_tiles, 3511 intptr_t st_off, int st_size, 3512 jlong con) { 3513 if ((st_off & (st_size-1)) != 0) 3514 return false; // strange store offset (assume size==2**N) 3515 address addr = (address)tiles + st_off; 3516 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); 3517 switch (st_size) { 3518 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; 3519 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; 3520 case sizeof(jint): *(jint*) addr = (jint) con; break; 3521 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; 3522 default: return false; // strange store size (detect size!=2**N here) 3523 } 3524 return true; // return success to caller 3525 } 3526 3527 // Coalesce subword constants into int constants and possibly 3528 // into long constants. The goal, if the CPU permits, 3529 // is to initialize the object with a small number of 64-bit tiles. 3530 // Also, convert floating-point constants to bit patterns. 3531 // Non-constants are not relevant to this pass. 3532 // 3533 // In terms of the running example on InitializeNode::InitializeNode 3534 // and InitializeNode::capture_store, here is the transformation 3535 // of rawstore1 and rawstore2 into rawstore12: 3536 // alloc = (Allocate ...) 3537 // rawoop = alloc.RawAddress 3538 // tile12 = 0x00010002 3539 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) 3540 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) 3541 // 3542 void 3543 InitializeNode::coalesce_subword_stores(intptr_t header_size, 3544 Node* size_in_bytes, 3545 PhaseGVN* phase) { 3546 Compile* C = phase->C; 3547 3548 assert(stores_are_sane(phase), ""); 3549 // Note: After this pass, they are not completely sane, 3550 // since there may be some overlaps. 3551 3552 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; 3553 3554 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); 3555 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); 3556 size_limit = MIN2(size_limit, ti_limit); 3557 size_limit = align_size_up(size_limit, BytesPerLong); 3558 int num_tiles = size_limit / BytesPerLong; 3559 3560 // allocate space for the tile map: 3561 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small 3562 jlong tiles_buf[small_len]; 3563 Node* nodes_buf[small_len]; 3564 jlong inits_buf[small_len]; 3565 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] 3566 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3567 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] 3568 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); 3569 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] 3570 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); 3571 // tiles: exact bitwise model of all primitive constants 3572 // nodes: last constant-storing node subsumed into the tiles model 3573 // inits: which bytes (in each tile) are touched by any initializations 3574 3575 //// Pass A: Fill in the tile model with any relevant stores. 3576 3577 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); 3578 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); 3579 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); 3580 Node* zmem = zero_memory(); // initially zero memory state 3581 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3582 Node* st = in(i); 3583 intptr_t st_off = get_store_offset(st, phase); 3584 3585 // Figure out the store's offset and constant value: 3586 if (st_off < header_size) continue; //skip (ignore header) 3587 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) 3588 int st_size = st->as_Store()->memory_size(); 3589 if (st_off + st_size > size_limit) break; 3590 3591 // Record which bytes are touched, whether by constant or not. 3592 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) 3593 continue; // skip (strange store size) 3594 3595 const Type* val = phase->type(st->in(MemNode::ValueIn)); 3596 if (!val->singleton()) continue; //skip (non-con store) 3597 BasicType type = val->basic_type(); 3598 3599 jlong con = 0; 3600 switch (type) { 3601 case T_INT: con = val->is_int()->get_con(); break; 3602 case T_LONG: con = val->is_long()->get_con(); break; 3603 case T_FLOAT: con = jint_cast(val->getf()); break; 3604 case T_DOUBLE: con = jlong_cast(val->getd()); break; 3605 default: continue; //skip (odd store type) 3606 } 3607 3608 if (type == T_LONG && Matcher::isSimpleConstant64(con) && 3609 st->Opcode() == Op_StoreL) { 3610 continue; // This StoreL is already optimal. 3611 } 3612 3613 // Store down the constant. 3614 store_constant(tiles, num_tiles, st_off, st_size, con); 3615 3616 intptr_t j = st_off >> LogBytesPerLong; 3617 3618 if (type == T_INT && st_size == BytesPerInt 3619 && (st_off & BytesPerInt) == BytesPerInt) { 3620 jlong lcon = tiles[j]; 3621 if (!Matcher::isSimpleConstant64(lcon) && 3622 st->Opcode() == Op_StoreI) { 3623 // This StoreI is already optimal by itself. 3624 jint* intcon = (jint*) &tiles[j]; 3625 intcon[1] = 0; // undo the store_constant() 3626 3627 // If the previous store is also optimal by itself, back up and 3628 // undo the action of the previous loop iteration... if we can. 3629 // But if we can't, just let the previous half take care of itself. 3630 st = nodes[j]; 3631 st_off -= BytesPerInt; 3632 con = intcon[0]; 3633 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { 3634 assert(st_off >= header_size, "still ignoring header"); 3635 assert(get_store_offset(st, phase) == st_off, "must be"); 3636 assert(in(i-1) == zmem, "must be"); 3637 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); 3638 assert(con == tcon->is_int()->get_con(), "must be"); 3639 // Undo the effects of the previous loop trip, which swallowed st: 3640 intcon[0] = 0; // undo store_constant() 3641 set_req(i-1, st); // undo set_req(i, zmem) 3642 nodes[j] = NULL; // undo nodes[j] = st 3643 --old_subword; // undo ++old_subword 3644 } 3645 continue; // This StoreI is already optimal. 3646 } 3647 } 3648 3649 // This store is not needed. 3650 set_req(i, zmem); 3651 nodes[j] = st; // record for the moment 3652 if (st_size < BytesPerLong) // something has changed 3653 ++old_subword; // includes int/float, but who's counting... 3654 else ++old_long; 3655 } 3656 3657 if ((old_subword + old_long) == 0) 3658 return; // nothing more to do 3659 3660 //// Pass B: Convert any non-zero tiles into optimal constant stores. 3661 // Be sure to insert them before overlapping non-constant stores. 3662 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) 3663 for (int j = 0; j < num_tiles; j++) { 3664 jlong con = tiles[j]; 3665 jlong init = inits[j]; 3666 if (con == 0) continue; 3667 jint con0, con1; // split the constant, address-wise 3668 jint init0, init1; // split the init map, address-wise 3669 { union { jlong con; jint intcon[2]; } u; 3670 u.con = con; 3671 con0 = u.intcon[0]; 3672 con1 = u.intcon[1]; 3673 u.con = init; 3674 init0 = u.intcon[0]; 3675 init1 = u.intcon[1]; 3676 } 3677 3678 Node* old = nodes[j]; 3679 assert(old != NULL, "need the prior store"); 3680 intptr_t offset = (j * BytesPerLong); 3681 3682 bool split = !Matcher::isSimpleConstant64(con); 3683 3684 if (offset < header_size) { 3685 assert(offset + BytesPerInt >= header_size, "second int counts"); 3686 assert(*(jint*)&tiles[j] == 0, "junk in header"); 3687 split = true; // only the second word counts 3688 // Example: int a[] = { 42 ... } 3689 } else if (con0 == 0 && init0 == -1) { 3690 split = true; // first word is covered by full inits 3691 // Example: int a[] = { ... foo(), 42 ... } 3692 } else if (con1 == 0 && init1 == -1) { 3693 split = true; // second word is covered by full inits 3694 // Example: int a[] = { ... 42, foo() ... } 3695 } 3696 3697 // Here's a case where init0 is neither 0 nor -1: 3698 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } 3699 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. 3700 // In this case the tile is not split; it is (jlong)42. 3701 // The big tile is stored down, and then the foo() value is inserted. 3702 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) 3703 3704 Node* ctl = old->in(MemNode::Control); 3705 Node* adr = make_raw_address(offset, phase); 3706 const TypePtr* atp = TypeRawPtr::BOTTOM; 3707 3708 // One or two coalesced stores to plop down. 3709 Node* st[2]; 3710 intptr_t off[2]; 3711 int nst = 0; 3712 if (!split) { 3713 ++new_long; 3714 off[nst] = offset; 3715 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3716 phase->longcon(con), T_LONG, MemNode::unordered); 3717 } else { 3718 // Omit either if it is a zero. 3719 if (con0 != 0) { 3720 ++new_int; 3721 off[nst] = offset; 3722 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3723 phase->intcon(con0), T_INT, MemNode::unordered); 3724 } 3725 if (con1 != 0) { 3726 ++new_int; 3727 offset += BytesPerInt; 3728 adr = make_raw_address(offset, phase); 3729 off[nst] = offset; 3730 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, 3731 phase->intcon(con1), T_INT, MemNode::unordered); 3732 } 3733 } 3734 3735 // Insert second store first, then the first before the second. 3736 // Insert each one just before any overlapping non-constant stores. 3737 while (nst > 0) { 3738 Node* st1 = st[--nst]; 3739 C->copy_node_notes_to(st1, old); 3740 st1 = phase->transform(st1); 3741 offset = off[nst]; 3742 assert(offset >= header_size, "do not smash header"); 3743 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); 3744 guarantee(ins_idx != 0, "must re-insert constant store"); 3745 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap 3746 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) 3747 set_req(--ins_idx, st1); 3748 else 3749 ins_req(ins_idx, st1); 3750 } 3751 } 3752 3753 if (PrintCompilation && WizardMode) 3754 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", 3755 old_subword, old_long, new_int, new_long); 3756 if (C->log() != NULL) 3757 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", 3758 old_subword, old_long, new_int, new_long); 3759 3760 // Clean up any remaining occurrences of zmem: 3761 remove_extra_zeroes(); 3762 } 3763 3764 // Explore forward from in(start) to find the first fully initialized 3765 // word, and return its offset. Skip groups of subword stores which 3766 // together initialize full words. If in(start) is itself part of a 3767 // fully initialized word, return the offset of in(start). If there 3768 // are no following full-word stores, or if something is fishy, return 3769 // a negative value. 3770 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { 3771 int int_map = 0; 3772 intptr_t int_map_off = 0; 3773 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for 3774 3775 for (uint i = start, limit = req(); i < limit; i++) { 3776 Node* st = in(i); 3777 3778 intptr_t st_off = get_store_offset(st, phase); 3779 if (st_off < 0) break; // return conservative answer 3780 3781 int st_size = st->as_Store()->memory_size(); 3782 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { 3783 return st_off; // we found a complete word init 3784 } 3785 3786 // update the map: 3787 3788 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); 3789 if (this_int_off != int_map_off) { 3790 // reset the map: 3791 int_map = 0; 3792 int_map_off = this_int_off; 3793 } 3794 3795 int subword_off = st_off - this_int_off; 3796 int_map |= right_n_bits(st_size) << subword_off; 3797 if ((int_map & FULL_MAP) == FULL_MAP) { 3798 return this_int_off; // we found a complete word init 3799 } 3800 3801 // Did this store hit or cross the word boundary? 3802 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); 3803 if (next_int_off == this_int_off + BytesPerInt) { 3804 // We passed the current int, without fully initializing it. 3805 int_map_off = next_int_off; 3806 int_map >>= BytesPerInt; 3807 } else if (next_int_off > this_int_off + BytesPerInt) { 3808 // We passed the current and next int. 3809 return this_int_off + BytesPerInt; 3810 } 3811 } 3812 3813 return -1; 3814 } 3815 3816 3817 // Called when the associated AllocateNode is expanded into CFG. 3818 // At this point, we may perform additional optimizations. 3819 // Linearize the stores by ascending offset, to make memory 3820 // activity as coherent as possible. 3821 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, 3822 intptr_t header_size, 3823 Node* size_in_bytes, 3824 PhaseGVN* phase) { 3825 assert(!is_complete(), "not already complete"); 3826 assert(stores_are_sane(phase), ""); 3827 assert(allocation() != NULL, "must be present"); 3828 3829 remove_extra_zeroes(); 3830 3831 if (ReduceFieldZeroing || ReduceBulkZeroing) 3832 // reduce instruction count for common initialization patterns 3833 coalesce_subword_stores(header_size, size_in_bytes, phase); 3834 3835 Node* zmem = zero_memory(); // initially zero memory state 3836 Node* inits = zmem; // accumulating a linearized chain of inits 3837 #ifdef ASSERT 3838 intptr_t first_offset = allocation()->minimum_header_size(); 3839 intptr_t last_init_off = first_offset; // previous init offset 3840 intptr_t last_init_end = first_offset; // previous init offset+size 3841 intptr_t last_tile_end = first_offset; // previous tile offset+size 3842 #endif 3843 intptr_t zeroes_done = header_size; 3844 3845 bool do_zeroing = true; // we might give up if inits are very sparse 3846 int big_init_gaps = 0; // how many large gaps have we seen? 3847 3848 if (ZeroTLAB) do_zeroing = false; 3849 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; 3850 3851 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { 3852 Node* st = in(i); 3853 intptr_t st_off = get_store_offset(st, phase); 3854 if (st_off < 0) 3855 break; // unknown junk in the inits 3856 if (st->in(MemNode::Memory) != zmem) 3857 break; // complicated store chains somehow in list 3858 3859 int st_size = st->as_Store()->memory_size(); 3860 intptr_t next_init_off = st_off + st_size; 3861 3862 if (do_zeroing && zeroes_done < next_init_off) { 3863 // See if this store needs a zero before it or under it. 3864 intptr_t zeroes_needed = st_off; 3865 3866 if (st_size < BytesPerInt) { 3867 // Look for subword stores which only partially initialize words. 3868 // If we find some, we must lay down some word-level zeroes first, 3869 // underneath the subword stores. 3870 // 3871 // Examples: 3872 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s 3873 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y 3874 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z 3875 // 3876 // Note: coalesce_subword_stores may have already done this, 3877 // if it was prompted by constant non-zero subword initializers. 3878 // But this case can still arise with non-constant stores. 3879 3880 intptr_t next_full_store = find_next_fullword_store(i, phase); 3881 3882 // In the examples above: 3883 // in(i) p q r s x y z 3884 // st_off 12 13 14 15 12 13 14 3885 // st_size 1 1 1 1 1 1 1 3886 // next_full_s. 12 16 16 16 16 16 16 3887 // z's_done 12 16 16 16 12 16 12 3888 // z's_needed 12 16 16 16 16 16 16 3889 // zsize 0 0 0 0 4 0 4 3890 if (next_full_store < 0) { 3891 // Conservative tack: Zero to end of current word. 3892 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); 3893 } else { 3894 // Zero to beginning of next fully initialized word. 3895 // Or, don't zero at all, if we are already in that word. 3896 assert(next_full_store >= zeroes_needed, "must go forward"); 3897 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); 3898 zeroes_needed = next_full_store; 3899 } 3900 } 3901 3902 if (zeroes_needed > zeroes_done) { 3903 intptr_t zsize = zeroes_needed - zeroes_done; 3904 // Do some incremental zeroing on rawmem, in parallel with inits. 3905 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3906 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3907 zeroes_done, zeroes_needed, 3908 phase); 3909 zeroes_done = zeroes_needed; 3910 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) 3911 do_zeroing = false; // leave the hole, next time 3912 } 3913 } 3914 3915 // Collect the store and move on: 3916 st->set_req(MemNode::Memory, inits); 3917 inits = st; // put it on the linearized chain 3918 set_req(i, zmem); // unhook from previous position 3919 3920 if (zeroes_done == st_off) 3921 zeroes_done = next_init_off; 3922 3923 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); 3924 3925 #ifdef ASSERT 3926 // Various order invariants. Weaker than stores_are_sane because 3927 // a large constant tile can be filled in by smaller non-constant stores. 3928 assert(st_off >= last_init_off, "inits do not reverse"); 3929 last_init_off = st_off; 3930 const Type* val = NULL; 3931 if (st_size >= BytesPerInt && 3932 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && 3933 (int)val->basic_type() < (int)T_OBJECT) { 3934 assert(st_off >= last_tile_end, "tiles do not overlap"); 3935 assert(st_off >= last_init_end, "tiles do not overwrite inits"); 3936 last_tile_end = MAX2(last_tile_end, next_init_off); 3937 } else { 3938 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); 3939 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); 3940 assert(st_off >= last_init_end, "inits do not overlap"); 3941 last_init_end = next_init_off; // it's a non-tile 3942 } 3943 #endif //ASSERT 3944 } 3945 3946 remove_extra_zeroes(); // clear out all the zmems left over 3947 add_req(inits); 3948 3949 if (!ZeroTLAB) { 3950 // If anything remains to be zeroed, zero it all now. 3951 zeroes_done = align_size_down(zeroes_done, BytesPerInt); 3952 // if it is the last unused 4 bytes of an instance, forget about it 3953 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); 3954 if (zeroes_done + BytesPerLong >= size_limit) { 3955 assert(allocation() != NULL, ""); 3956 if (allocation()->Opcode() == Op_Allocate) { 3957 Node* klass_node = allocation()->in(AllocateNode::KlassNode); 3958 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); 3959 if (zeroes_done == k->layout_helper()) 3960 zeroes_done = size_limit; 3961 } 3962 } 3963 if (zeroes_done < size_limit) { 3964 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, 3965 zeroes_done, size_in_bytes, phase); 3966 } 3967 } 3968 3969 set_complete(phase); 3970 return rawmem; 3971 } 3972 3973 3974 #ifdef ASSERT 3975 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { 3976 if (is_complete()) 3977 return true; // stores could be anything at this point 3978 assert(allocation() != NULL, "must be present"); 3979 intptr_t last_off = allocation()->minimum_header_size(); 3980 for (uint i = InitializeNode::RawStores; i < req(); i++) { 3981 Node* st = in(i); 3982 intptr_t st_off = get_store_offset(st, phase); 3983 if (st_off < 0) continue; // ignore dead garbage 3984 if (last_off > st_off) { 3985 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off); 3986 this->dump(2); 3987 assert(false, "ascending store offsets"); 3988 return false; 3989 } 3990 last_off = st_off + st->as_Store()->memory_size(); 3991 } 3992 return true; 3993 } 3994 #endif //ASSERT 3995 3996 3997 3998 3999 //============================MergeMemNode===================================== 4000 // 4001 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several 4002 // contributing store or call operations. Each contributor provides the memory 4003 // state for a particular "alias type" (see Compile::alias_type). For example, 4004 // if a MergeMem has an input X for alias category #6, then any memory reference 4005 // to alias category #6 may use X as its memory state input, as an exact equivalent 4006 // to using the MergeMem as a whole. 4007 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) 4008 // 4009 // (Here, the <N> notation gives the index of the relevant adr_type.) 4010 // 4011 // In one special case (and more cases in the future), alias categories overlap. 4012 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory 4013 // states. Therefore, if a MergeMem has only one contributing input W for Bot, 4014 // it is exactly equivalent to that state W: 4015 // MergeMem(<Bot>: W) <==> W 4016 // 4017 // Usually, the merge has more than one input. In that case, where inputs 4018 // overlap (i.e., one is Bot), the narrower alias type determines the memory 4019 // state for that type, and the wider alias type (Bot) fills in everywhere else: 4020 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) 4021 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) 4022 // 4023 // A merge can take a "wide" memory state as one of its narrow inputs. 4024 // This simply means that the merge observes out only the relevant parts of 4025 // the wide input. That is, wide memory states arriving at narrow merge inputs 4026 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) 4027 // 4028 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), 4029 // and that memory slices "leak through": 4030 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) 4031 // 4032 // But, in such a cascade, repeated memory slices can "block the leak": 4033 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') 4034 // 4035 // In the last example, Y is not part of the combined memory state of the 4036 // outermost MergeMem. The system must, of course, prevent unschedulable 4037 // memory states from arising, so you can be sure that the state Y is somehow 4038 // a precursor to state Y'. 4039 // 4040 // 4041 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array 4042 // of each MergeMemNode array are exactly the numerical alias indexes, including 4043 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions 4044 // Compile::alias_type (and kin) produce and manage these indexes. 4045 // 4046 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. 4047 // (Note that this provides quick access to the top node inside MergeMem methods, 4048 // without the need to reach out via TLS to Compile::current.) 4049 // 4050 // As a consequence of what was just described, a MergeMem that represents a full 4051 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, 4052 // containing all alias categories. 4053 // 4054 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. 4055 // 4056 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either 4057 // a memory state for the alias type <N>, or else the top node, meaning that 4058 // there is no particular input for that alias type. Note that the length of 4059 // a MergeMem is variable, and may be extended at any time to accommodate new 4060 // memory states at larger alias indexes. When merges grow, they are of course 4061 // filled with "top" in the unused in() positions. 4062 // 4063 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. 4064 // (Top was chosen because it works smoothly with passes like GCM.) 4065 // 4066 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is 4067 // the type of random VM bits like TLS references.) Since it is always the 4068 // first non-Bot memory slice, some low-level loops use it to initialize an 4069 // index variable: for (i = AliasIdxRaw; i < req(); i++). 4070 // 4071 // 4072 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns 4073 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns 4074 // the memory state for alias type <N>, or (if there is no particular slice at <N>, 4075 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> 4076 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over 4077 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. 4078 // 4079 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't 4080 // really that different from the other memory inputs. An abbreviation called 4081 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. 4082 // 4083 // 4084 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent 4085 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi 4086 // that "emerges though" the base memory will be marked as excluding the alias types 4087 // of the other (narrow-memory) copies which "emerged through" the narrow edges: 4088 // 4089 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) 4090 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) 4091 // 4092 // This strange "subtraction" effect is necessary to ensure IGVN convergence. 4093 // (It is currently unimplemented.) As you can see, the resulting merge is 4094 // actually a disjoint union of memory states, rather than an overlay. 4095 // 4096 4097 //------------------------------MergeMemNode----------------------------------- 4098 Node* MergeMemNode::make_empty_memory() { 4099 Node* empty_memory = (Node*) Compile::current()->top(); 4100 assert(empty_memory->is_top(), "correct sentinel identity"); 4101 return empty_memory; 4102 } 4103 4104 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { 4105 init_class_id(Class_MergeMem); 4106 // all inputs are nullified in Node::Node(int) 4107 // set_input(0, NULL); // no control input 4108 4109 // Initialize the edges uniformly to top, for starters. 4110 Node* empty_mem = make_empty_memory(); 4111 for (uint i = Compile::AliasIdxTop; i < req(); i++) { 4112 init_req(i,empty_mem); 4113 } 4114 assert(empty_memory() == empty_mem, ""); 4115 4116 if( new_base != NULL && new_base->is_MergeMem() ) { 4117 MergeMemNode* mdef = new_base->as_MergeMem(); 4118 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); 4119 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { 4120 mms.set_memory(mms.memory2()); 4121 } 4122 assert(base_memory() == mdef->base_memory(), ""); 4123 } else { 4124 set_base_memory(new_base); 4125 } 4126 } 4127 4128 // Make a new, untransformed MergeMem with the same base as 'mem'. 4129 // If mem is itself a MergeMem, populate the result with the same edges. 4130 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { 4131 return new MergeMemNode(mem); 4132 } 4133 4134 //------------------------------cmp-------------------------------------------- 4135 uint MergeMemNode::hash() const { return NO_HASH; } 4136 uint MergeMemNode::cmp( const Node &n ) const { 4137 return (&n == this); // Always fail except on self 4138 } 4139 4140 //------------------------------Identity--------------------------------------- 4141 Node* MergeMemNode::Identity(PhaseTransform *phase) { 4142 // Identity if this merge point does not record any interesting memory 4143 // disambiguations. 4144 Node* base_mem = base_memory(); 4145 Node* empty_mem = empty_memory(); 4146 if (base_mem != empty_mem) { // Memory path is not dead? 4147 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4148 Node* mem = in(i); 4149 if (mem != empty_mem && mem != base_mem) { 4150 return this; // Many memory splits; no change 4151 } 4152 } 4153 } 4154 return base_mem; // No memory splits; ID on the one true input 4155 } 4156 4157 //------------------------------Ideal------------------------------------------ 4158 // This method is invoked recursively on chains of MergeMem nodes 4159 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { 4160 // Remove chain'd MergeMems 4161 // 4162 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted 4163 // relative to the "in(Bot)". Since we are patching both at the same time, 4164 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", 4165 // but rewrite each "in(i)" relative to the new "in(Bot)". 4166 Node *progress = NULL; 4167 4168 4169 Node* old_base = base_memory(); 4170 Node* empty_mem = empty_memory(); 4171 if (old_base == empty_mem) 4172 return NULL; // Dead memory path. 4173 4174 MergeMemNode* old_mbase; 4175 if (old_base != NULL && old_base->is_MergeMem()) 4176 old_mbase = old_base->as_MergeMem(); 4177 else 4178 old_mbase = NULL; 4179 Node* new_base = old_base; 4180 4181 // simplify stacked MergeMems in base memory 4182 if (old_mbase) new_base = old_mbase->base_memory(); 4183 4184 // the base memory might contribute new slices beyond my req() 4185 if (old_mbase) grow_to_match(old_mbase); 4186 4187 // Look carefully at the base node if it is a phi. 4188 PhiNode* phi_base; 4189 if (new_base != NULL && new_base->is_Phi()) 4190 phi_base = new_base->as_Phi(); 4191 else 4192 phi_base = NULL; 4193 4194 Node* phi_reg = NULL; 4195 uint phi_len = (uint)-1; 4196 if (phi_base != NULL && !phi_base->is_copy()) { 4197 // do not examine phi if degraded to a copy 4198 phi_reg = phi_base->region(); 4199 phi_len = phi_base->req(); 4200 // see if the phi is unfinished 4201 for (uint i = 1; i < phi_len; i++) { 4202 if (phi_base->in(i) == NULL) { 4203 // incomplete phi; do not look at it yet! 4204 phi_reg = NULL; 4205 phi_len = (uint)-1; 4206 break; 4207 } 4208 } 4209 } 4210 4211 // Note: We do not call verify_sparse on entry, because inputs 4212 // can normalize to the base_memory via subsume_node or similar 4213 // mechanisms. This method repairs that damage. 4214 4215 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); 4216 4217 // Look at each slice. 4218 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4219 Node* old_in = in(i); 4220 // calculate the old memory value 4221 Node* old_mem = old_in; 4222 if (old_mem == empty_mem) old_mem = old_base; 4223 assert(old_mem == memory_at(i), ""); 4224 4225 // maybe update (reslice) the old memory value 4226 4227 // simplify stacked MergeMems 4228 Node* new_mem = old_mem; 4229 MergeMemNode* old_mmem; 4230 if (old_mem != NULL && old_mem->is_MergeMem()) 4231 old_mmem = old_mem->as_MergeMem(); 4232 else 4233 old_mmem = NULL; 4234 if (old_mmem == this) { 4235 // This can happen if loops break up and safepoints disappear. 4236 // A merge of BotPtr (default) with a RawPtr memory derived from a 4237 // safepoint can be rewritten to a merge of the same BotPtr with 4238 // the BotPtr phi coming into the loop. If that phi disappears 4239 // also, we can end up with a self-loop of the mergemem. 4240 // In general, if loops degenerate and memory effects disappear, 4241 // a mergemem can be left looking at itself. This simply means 4242 // that the mergemem's default should be used, since there is 4243 // no longer any apparent effect on this slice. 4244 // Note: If a memory slice is a MergeMem cycle, it is unreachable 4245 // from start. Update the input to TOP. 4246 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; 4247 } 4248 else if (old_mmem != NULL) { 4249 new_mem = old_mmem->memory_at(i); 4250 } 4251 // else preceding memory was not a MergeMem 4252 4253 // replace equivalent phis (unfortunately, they do not GVN together) 4254 if (new_mem != NULL && new_mem != new_base && 4255 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { 4256 if (new_mem->is_Phi()) { 4257 PhiNode* phi_mem = new_mem->as_Phi(); 4258 for (uint i = 1; i < phi_len; i++) { 4259 if (phi_base->in(i) != phi_mem->in(i)) { 4260 phi_mem = NULL; 4261 break; 4262 } 4263 } 4264 if (phi_mem != NULL) { 4265 // equivalent phi nodes; revert to the def 4266 new_mem = new_base; 4267 } 4268 } 4269 } 4270 4271 // maybe store down a new value 4272 Node* new_in = new_mem; 4273 if (new_in == new_base) new_in = empty_mem; 4274 4275 if (new_in != old_in) { 4276 // Warning: Do not combine this "if" with the previous "if" 4277 // A memory slice might have be be rewritten even if it is semantically 4278 // unchanged, if the base_memory value has changed. 4279 set_req(i, new_in); 4280 progress = this; // Report progress 4281 } 4282 } 4283 4284 if (new_base != old_base) { 4285 set_req(Compile::AliasIdxBot, new_base); 4286 // Don't use set_base_memory(new_base), because we need to update du. 4287 assert(base_memory() == new_base, ""); 4288 progress = this; 4289 } 4290 4291 if( base_memory() == this ) { 4292 // a self cycle indicates this memory path is dead 4293 set_req(Compile::AliasIdxBot, empty_mem); 4294 } 4295 4296 // Resolve external cycles by calling Ideal on a MergeMem base_memory 4297 // Recursion must occur after the self cycle check above 4298 if( base_memory()->is_MergeMem() ) { 4299 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); 4300 Node *m = phase->transform(new_mbase); // Rollup any cycles 4301 if( m != NULL && (m->is_top() || 4302 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { 4303 // propagate rollup of dead cycle to self 4304 set_req(Compile::AliasIdxBot, empty_mem); 4305 } 4306 } 4307 4308 if( base_memory() == empty_mem ) { 4309 progress = this; 4310 // Cut inputs during Parse phase only. 4311 // During Optimize phase a dead MergeMem node will be subsumed by Top. 4312 if( !can_reshape ) { 4313 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4314 if( in(i) != empty_mem ) { set_req(i, empty_mem); } 4315 } 4316 } 4317 } 4318 4319 if( !progress && base_memory()->is_Phi() && can_reshape ) { 4320 // Check if PhiNode::Ideal's "Split phis through memory merges" 4321 // transform should be attempted. Look for this->phi->this cycle. 4322 uint merge_width = req(); 4323 if (merge_width > Compile::AliasIdxRaw) { 4324 PhiNode* phi = base_memory()->as_Phi(); 4325 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in 4326 if (phi->in(i) == this) { 4327 phase->is_IterGVN()->_worklist.push(phi); 4328 break; 4329 } 4330 } 4331 } 4332 } 4333 4334 assert(progress || verify_sparse(), "please, no dups of base"); 4335 return progress; 4336 } 4337 4338 //-------------------------set_base_memory------------------------------------- 4339 void MergeMemNode::set_base_memory(Node *new_base) { 4340 Node* empty_mem = empty_memory(); 4341 set_req(Compile::AliasIdxBot, new_base); 4342 assert(memory_at(req()) == new_base, "must set default memory"); 4343 // Clear out other occurrences of new_base: 4344 if (new_base != empty_mem) { 4345 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4346 if (in(i) == new_base) set_req(i, empty_mem); 4347 } 4348 } 4349 } 4350 4351 //------------------------------out_RegMask------------------------------------ 4352 const RegMask &MergeMemNode::out_RegMask() const { 4353 return RegMask::Empty; 4354 } 4355 4356 //------------------------------dump_spec-------------------------------------- 4357 #ifndef PRODUCT 4358 void MergeMemNode::dump_spec(outputStream *st) const { 4359 st->print(" {"); 4360 Node* base_mem = base_memory(); 4361 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { 4362 Node* mem = memory_at(i); 4363 if (mem == base_mem) { st->print(" -"); continue; } 4364 st->print( " N%d:", mem->_idx ); 4365 Compile::current()->get_adr_type(i)->dump_on(st); 4366 } 4367 st->print(" }"); 4368 } 4369 #endif // !PRODUCT 4370 4371 4372 #ifdef ASSERT 4373 static bool might_be_same(Node* a, Node* b) { 4374 if (a == b) return true; 4375 if (!(a->is_Phi() || b->is_Phi())) return false; 4376 // phis shift around during optimization 4377 return true; // pretty stupid... 4378 } 4379 4380 // verify a narrow slice (either incoming or outgoing) 4381 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { 4382 if (!VerifyAliases) return; // don't bother to verify unless requested 4383 if (is_error_reported()) return; // muzzle asserts when debugging an error 4384 if (Node::in_dump()) return; // muzzle asserts when printing 4385 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); 4386 assert(n != NULL, ""); 4387 // Elide intervening MergeMem's 4388 while (n->is_MergeMem()) { 4389 n = n->as_MergeMem()->memory_at(alias_idx); 4390 } 4391 Compile* C = Compile::current(); 4392 const TypePtr* n_adr_type = n->adr_type(); 4393 if (n == m->empty_memory()) { 4394 // Implicit copy of base_memory() 4395 } else if (n_adr_type != TypePtr::BOTTOM) { 4396 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); 4397 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); 4398 } else { 4399 // A few places like make_runtime_call "know" that VM calls are narrow, 4400 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. 4401 bool expected_wide_mem = false; 4402 if (n == m->base_memory()) { 4403 expected_wide_mem = true; 4404 } else if (alias_idx == Compile::AliasIdxRaw || 4405 n == m->memory_at(Compile::AliasIdxRaw)) { 4406 expected_wide_mem = true; 4407 } else if (!C->alias_type(alias_idx)->is_rewritable()) { 4408 // memory can "leak through" calls on channels that 4409 // are write-once. Allow this also. 4410 expected_wide_mem = true; 4411 } 4412 assert(expected_wide_mem, "expected narrow slice replacement"); 4413 } 4414 } 4415 #else // !ASSERT 4416 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op 4417 #endif 4418 4419 4420 //-----------------------------memory_at--------------------------------------- 4421 Node* MergeMemNode::memory_at(uint alias_idx) const { 4422 assert(alias_idx >= Compile::AliasIdxRaw || 4423 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, 4424 "must avoid base_memory and AliasIdxTop"); 4425 4426 // Otherwise, it is a narrow slice. 4427 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); 4428 Compile *C = Compile::current(); 4429 if (is_empty_memory(n)) { 4430 // the array is sparse; empty slots are the "top" node 4431 n = base_memory(); 4432 assert(Node::in_dump() 4433 || n == NULL || n->bottom_type() == Type::TOP 4434 || n->adr_type() == NULL // address is TOP 4435 || n->adr_type() == TypePtr::BOTTOM 4436 || n->adr_type() == TypeRawPtr::BOTTOM 4437 || Compile::current()->AliasLevel() == 0, 4438 "must be a wide memory"); 4439 // AliasLevel == 0 if we are organizing the memory states manually. 4440 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. 4441 } else { 4442 // make sure the stored slice is sane 4443 #ifdef ASSERT 4444 if (is_error_reported() || Node::in_dump()) { 4445 } else if (might_be_same(n, base_memory())) { 4446 // Give it a pass: It is a mostly harmless repetition of the base. 4447 // This can arise normally from node subsumption during optimization. 4448 } else { 4449 verify_memory_slice(this, alias_idx, n); 4450 } 4451 #endif 4452 } 4453 return n; 4454 } 4455 4456 //---------------------------set_memory_at------------------------------------- 4457 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { 4458 verify_memory_slice(this, alias_idx, n); 4459 Node* empty_mem = empty_memory(); 4460 if (n == base_memory()) n = empty_mem; // collapse default 4461 uint need_req = alias_idx+1; 4462 if (req() < need_req) { 4463 if (n == empty_mem) return; // already the default, so do not grow me 4464 // grow the sparse array 4465 do { 4466 add_req(empty_mem); 4467 } while (req() < need_req); 4468 } 4469 set_req( alias_idx, n ); 4470 } 4471 4472 4473 4474 //--------------------------iteration_setup------------------------------------ 4475 void MergeMemNode::iteration_setup(const MergeMemNode* other) { 4476 if (other != NULL) { 4477 grow_to_match(other); 4478 // invariant: the finite support of mm2 is within mm->req() 4479 #ifdef ASSERT 4480 for (uint i = req(); i < other->req(); i++) { 4481 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); 4482 } 4483 #endif 4484 } 4485 // Replace spurious copies of base_memory by top. 4486 Node* base_mem = base_memory(); 4487 if (base_mem != NULL && !base_mem->is_top()) { 4488 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { 4489 if (in(i) == base_mem) 4490 set_req(i, empty_memory()); 4491 } 4492 } 4493 } 4494 4495 //---------------------------grow_to_match------------------------------------- 4496 void MergeMemNode::grow_to_match(const MergeMemNode* other) { 4497 Node* empty_mem = empty_memory(); 4498 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); 4499 // look for the finite support of the other memory 4500 for (uint i = other->req(); --i >= req(); ) { 4501 if (other->in(i) != empty_mem) { 4502 uint new_len = i+1; 4503 while (req() < new_len) add_req(empty_mem); 4504 break; 4505 } 4506 } 4507 } 4508 4509 //---------------------------verify_sparse------------------------------------- 4510 #ifndef PRODUCT 4511 bool MergeMemNode::verify_sparse() const { 4512 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); 4513 Node* base_mem = base_memory(); 4514 // The following can happen in degenerate cases, since empty==top. 4515 if (is_empty_memory(base_mem)) return true; 4516 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { 4517 assert(in(i) != NULL, "sane slice"); 4518 if (in(i) == base_mem) return false; // should have been the sentinel value! 4519 } 4520 return true; 4521 } 4522 4523 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { 4524 Node* n; 4525 n = mm->in(idx); 4526 if (mem == n) return true; // might be empty_memory() 4527 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); 4528 if (mem == n) return true; 4529 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { 4530 if (mem == n) return true; 4531 if (n == NULL) break; 4532 } 4533 return false; 4534 } 4535 #endif // !PRODUCT